Collaboration Networks of the Implementation Science Centers for Cancer Control (ISC3): A Social Network Analysis

Rebekah R Jacob (  rebekahjacob@wustl.edu ) Washington University in Saint Louis https://orcid.org/0000-0003-2466-8809 Ariella R Korn National Cancer Institute Grace C Huang Westat Douglas Easterling Wake Forest University School of Medicine Daniel A Gundersen Dana-Farber Cancer Institute Shoba Ramanadhan Harvard University T H Chan School of Public Health Thuy Vu University of Washington School of Public Health Heather Angier Oregon Health & Science University Ross C Brownson Washington University in St Louis Debra Haire-Joshu Washington University in St Louis April Oh National Cancer Institute Robert Schnoll University of Pennsylvania


Background
In the United States, an estimated 1.9 million people are diagnosed with cancer each year (1,2). While cancer mortality rates have decreased overall from 2001 to 2018, they have done so disproportionately in magnitude across cancer types and certain racial/ethnic groups (1,3). In 2030, it is estimated that the US will spend $246 billion on cancer-related healthcare, a 34% increase from 2015 (4). Effective interventions exist to reduce cancer burden, but they are largely dependent on ne-tuned implementation at the policy and practice level (health care and public health settings), which often lags approximately 17 years behind the research (5,6). Discovering the tools and methods to close this "know-do" gap is the charge of the Implementation Science (IS) eld (7,8) and is a priority in cancer research initiatives and funding.
In 2018, the National Cancer Institute (NCI) issued requests for funding to support Implementation Science Centers for Cancer Control (ISC 3 ) (2019-2024) (9)(10)(11). Each of the seven funded ISC 3 centers conducts pilot work in IS approaches and includes an "implementation laboratory" that provides important linkages with clinical and community sites. Through this infrastructure, the ISC 3 aims to build scienti c capacity in the eld with targeted approaches for developing and testing innovative methods and measures for dissemination and implementation research and engaging scholars in a rich network of investigators (11). ISC 3 has the potential to foster eld-wide collaboration with multiple funded centers addressing important and timely IS research in cancer control. Collaboration among ISC 3 investigators and staff within and across centers is critical and can lead to greater productivity and impact, diverse thinking, and increased opportunities for capacity building in the eld (12)(13)(14). Priming the network to develop additional scienti c linkages between researchers is a key focus of the ISC 3 , and therefore, understanding the extent of these connections is an important evaluation priority.
In addition to understanding the science generated and disseminated among network members (publications, presentations, funding), social network methods can be utilized to establish a baseline assessment of collaboration among investigators and identify where efforts should be allocated to achieve goals. Multi-organization initiatives and research networks have demonstrated the utility of social network measures for evaluation and for midway planning to enhance network cohesion (15)(16)(17)(18)(19)(20)(21). For example, the Translational Research Network conducted longitudinal surveys over four years that highlighted new researcherclinician linkages that occurred due to the network as well as projects not directly funded by the Translational Research Network but attributed to network membership (20). In addition, Vacca et al. used social network analysis to identify and match researchers with funding collaborations across various research communities (16,22). Understanding linkages within networks informs areas where growth or additional relationship types are desired and carried out through purposefully designed network interventions (23,24). This is especially important at the beginning of an initiative given the time to implement and measure outcomes resulting from the creation of new collaborations or other changes to linkages (24). Likewise, these measures are key evaluation pieces for initiatives that seek to understand how efforts have impacted network functions and how collaboration aided productivity and created potentially durable infrastructure for further action.
The purpose of this study is to understand the scienti c collaborations and early linkages within the ISC 3 network during the initiative's rst year. We examined collaboration patterns overall and by distinct activities: planning or conducting research, capacity building, product development, scienti c dissemination, and practice/policy dissemination. In addition to serving as baseline, year 1 data will be used to inform additional network reassessments planned for years 3 and 5. This research is expected to inform ISC 3 's efforts to strengthen ties and support the development of new IS collaborations to bridge speci c areas of IS and cancer control research.

Center descriptions
The seven funded centers that make up the ISC 3 are: All centers were funded starting in October 2019 except for Penn ISC 3 who entered as a new center in October 2020. Additional information on each center can be found at https://cancercontrol.cancer.gov/is/initiatives/isc3.

Cross-Center Evaluation Work Group
In the fall of 2019, a work group formed to develop cross-center evaluation measures with representation from each of the centers, NCI program staff, and by the initiative's contracted evaluator, Westat. The Cross-Center Evaluation work group developed a survey tool to assess both intra-and inter-center research collaborations that were of interest to the ISC 3 evaluation. The work group continues to meet monthly and is facilitated by center members.

Participants
Each center's leadership team assembled and provided a list of key researchers-faculty, staff, trainees, and others who were critical to their scope of work. The Cross-Center Evaluation work group developed inclusion criteria to assist centers in developing their network survey participant list. Inclusion criteria for ISC 3 network membership were: faculty members supported directly by ISC 3 funds and/or directly involved in ISC 3 activities; staff members supported by ISC 3 funds whose main role was research (research staff); and doctoral and post-doctoral trainees supported directly through ISC 3 funds and/or who were involved in the center's research. NCI ISC 3 program staff and leadership were also included as invitees because of their role in many of the network activities, including product development and capacity building. Across the seven ISC 3 centers and NCI, a total of 192 individuals were invited to participate (range 11-51 invitees per center).

Data collection
Westat evaluators invited participants by email to complete a 10-25 minute web-based network survey in September 2020. The Cross-Center Evaluation work group provided input on the email invitation text and center leadership were asked to send an introductory email to precede the invitation. NCI leadership presented information about the annual ISC 3 meeting in September 2020 and encouraged center members to participate. The online module included a feature for participants to save their survey progress and then return to nish at a later time. Automated reminder emails were delivered two times. In addition, center leaders were supplied weekly with their center's list of non-respondents.

Measures
The survey asked participants to identify those with whom they had direct contact within the past 12 months across a roster of all 192 invited individuals. The full survey is provided in Additional File 1. Examples of direct contact included meetings, work groups, phone calls, and emails (participants were asked not to count contact based solely upon listservs or mass emails). We measured contact in three categories: 1) "I do not know this person"; 2) "I know who this person is, but we have had no contact"; and 3) "I have had contact with this person at least once in the last 12 months." A set of follow-up questions asked about ve potential collaborative activities for individuals with whom the participant had contact in the past 12 months: 1) "Planned or conducted research (e.g., grant writing, study design or execution)" (hereafter referred to as planning/conducting research); 2) "Engaged in capacity building (e.g., trainings, learning communities, mentoring)" (hereafter referred to as capacity building); 3) "Developed products in cross-center work group or committee (e.g., measures database, survey instrument)" (hereafter referred to as product development); 4) "Disseminated research to a science audience (e.g., scholarly publication, conference presentation)" (hereafter referred to as scienti c dissemination); and 5) "Disseminated research to a nonscience audience (e.g., evaluation report, policy brief)" (hereafter referred to as practice/policy dissemination).
The survey also assessed various participant characteristics. Participants identi ed their scienti c discipline (public health, medicine, psychology, social work, other), length of years worked in their eld (less than 5 years, 5 to 9 years, 10-15 years, more than 15 years), role within the ISC 3 initiative (doctoral student, post-doc, staff, faculty, NCI staff, other), gender identity (male, female, transgender, gender non-conforming, other), and racial and ethnic background (American Indian or Alaska Native, Asian, Black or African American, Hispanic or Latino, Native Hawaiian or Paci c islander, white, other). Participants also identi ed their level of expertise in IS (beginner, intermediate, advanced).

Data Analysis
Network data were cleaned and analyzed using R with the igraph package for network analysis (25). Ties were symmetrized for undirected analysis, a common approach for networks where relational direction is not a major focus and collaboration is assumed from either direction (18, 26). Therefore, in our network, a nomination indicates a collaboration activity exists between two nodes and is undirected (if A nominates B, then A/B are connected whether or not B reciprocates the nomination). We explored the network structure visually (graphs) and descriptively across the ve separate activity networks and the network of all activities combined. Quantitative descriptive measures include density, degree centralization, betweenness centralization, transitivity, number of isolates, and proportion of collaborations within and across centers.
Density is a network measure that represents the overall connectivity or degree of cohesion within a network. It is calculated as the ratio of the number of ties to the total number of possible ties in the network, ranges from 0 to 1 and can be expressed as a proportion (27). Centralization measures hierarchy in networks or the extent that connections in the network are dependent on a select few most central nodes in the network (0 to 1). Degree centralization is calculated based on the number of connections where a higher degree centralization represents more hierarchy in the network or that one or more nodes hold most of the connections in the network (0 to 1). Betweenness centralization is calculated based on geodesic, or shortest paths, between nodes and is used to measure the extent to which networks contain bridge or gatekeepers where higher betweenness centralization would signify that one or a few nodes are responsible for holding the network together (0 to 1) (28). Transitivity is a measure of probability of triangles within a network or how likely ties are to form between nodes that share a common collaborator (0 to 1). For example, if A and B are connected, and B and C are connected, transitivity represents the probability that A and C are also connected (28). Isolates, or those nodes without any network connections, identify where opportunities for collaboration exist. We examined isolates for node-attribute level patterns. To measure the amount of collaboration happening between the different centers and NCI, we looked at percentage of all ties that were between members from separate centers.
For the number of nominations for each node, we examined median degree and range of ties. We used Kruskal-Wallis chi-square tests (non-parametric alternative to one-way ANOVA) to determine rank order differences in number of connections across categories of participant characteristics.
We conducted a sensitivity analysis to explore the in uence of extreme outdegree levels on our ndings (i.e., accounting for respondents who nominated many collaborators). To do so, we replicated all analyses using a reduced dataset that removed outgoing unreciprocated ties from "outlier" respondents. Outliers were de ned as those with an outdegree more than 1.5 times the interquartile range above the third quartile (Q3+1.5*IQR) for each type of collaboration. We used the STROBE cross-sectional reporting guidelines detailed in Additional File 2 (29).

Results
A total of 182 participants completed the survey (95% response rate) ranging from 91% to 100% across centers. In network analysis, survey non-respondents are still considered as members in the full network because respondents are still able to nominate non-respondents. Therefore, our total network contained 192 members. Characteristics of ISC 3 network members are reported in Table 1. Non-respondents (n=10) were excluded from counts/totals where participant characteristic information is necessary. Most participants reported their primary discipline as public health (54.4%) or medicine (24.2%). More than two-thirds (69.8%) of network members reported ten or more years of experience in their eld. The most prevalent network roles were faculty (60.4%) and center staff (23.6%). The largest proportion of network members reported intermediate expertise in IS (41.8%), followed by beginner (35.2%) and advanced (23.1%) expertise. Most members identi ed as female (68.7%) and white (78.7%).

Network Characteristics
With all collaboration activities combined, the ISC 3 network included 192 members with a total of 2480 collaboration ties, of which members had a median 22 connections (Table 2). Figure 1 shows the network for all collaboration activities combined and Figure  2 displays the network for each separate collaboration activity. The greatest number of ties were reported in planning/conducting research (1470 ties; median 15 ties/member) and the fewest ties were reported in practice/policy dissemination (284 ties; median 2 ties/member). The ISC 3 density for all collaboration activities was 13.5%. Across the different collaboration types, the most and least densely connected network activities were planning/conducting research (8.2%) and practice/policy dissemination (2.6%), respectively. The practice/policy dissemination network was the smallest network with just 143 of the 192 ISC 3 network members represented whereas the other networks ranged from 173-190 members.
The overall ISC 3 network was fairly decentralized (degree centralization=0.33 and betweenness centralization=0.07; Table 2) consistent with Figure 1's basic linked local networks shape (no strong central node or group of nodes). For separate activities, capacity building and product development had the highest degree centralization (0.23 and 0.21, respectively) compared to other collaboration activities, which ranged from 0.12 to 0.17, suggesting in uential positions for some members in these networks ("hub and spoke" network structure). Scienti c dissemination and practice/policy dissemination networks had the highest betweenness centralization (0.23 and 0.20, respectively), suggesting some members may be closer to each other and/or are more easily connected or reached. As more connections "pass" through these central members, their removal would result in high number of isolates.
Overall, the ISC 3 's transitivity (0.47) suggests heightened probability of triangles in the network, though variation exists across collaboration types. Planning/conducting research had the highest transitivity measure (0.56) compared to all other collaboration networks (transitivity range: 0.33 to 0.37), suggesting that two investigators that are collaborating with the same investigator are likely to also be collaborating with each other.
One-third of all collaboration ties (33.0%) occurred between members from different centers. For speci c collaboration activities, we observed the largest portion of cross-center collaboration in product development (48.1%), which includes involvement with cross-center work groups. Collaborating on practice/policy dissemination and planning/conducting research mostly occurred within members' respective centers (6.0% and 11.7% cross-center ties, respectively). Network members had a median 17 connections within their own center and 7 connections from other centers across all activities.
There were no isolates for the all collaboration activities network because our overall network was derived from having at least one collaboration activity reported. Notably, practice/policy dissemination and product development were the two activity networks with the largest number of isolates, 22% and 10% of the total number of network members (n=43 and n=19, respectively).
Descriptive analysis on isolates from these two activity networks showed that half of ISC 3 trainees (n=6) were not connected in product development whereas other roles represented 0-33% of isolates. For practice/policy dissemination, 44.4% (n=8) of those with less than 5 years of experience in their eld were not collaborating, compared to those categories with more experience (range 16.1%-16.9%). NCI staff (n=5) and trainees (n=4) also made up larger portions of isolates in the policy/practice dissemination network (45.5% and 33.3%, respectively).

Sensitivity Analysis
Full results from our sensitivity analysis are available in Additional File 3, in which we adjusted for extreme outdegree levels. For example, for all collaboration activities combined, the number of ties was reduced from 2480 to 2419 as a result of removing outgoing unreciprocated ties from "outlier" respondents (n=4) with an outdegree more than Q3+1.5*IQR. Network-and node-level characteristics were generally similar across the full and reduced datasets for each type of collaboration, however we note the

Discussion
Overall, the ISC 3 network in its rst year was highly connected within each center (intra) and had modest linkages between members from different centers (inter). We found that one-third of all ties were inter-center collaborations. Okamoto et al. (2015) demonstrated a similar measure of cross-center collaboration in their social network study across ten geographically distanced NCI Centers for Population Health and Health Disparities (CPHHD). CPHHD's ndings reported 7% cross-center ties (17). Though direct comparisons across networks should be made with caution, Okamoto et al.'s methods and ndings provide a similar approach to understanding linkages in a multi-center initiative. The infrastructure and time to develop and support a network-wide survey is considerable. This is likely re ected in the larger number of cross-center ties in CPHHD's prior funding 5-year funding cycle (17%).
We surveyed ISC 3 members at the end of our rst year of funding, likely in ating our cross-center ties to some degree. ISC 3 work groups had already begun to form and work on projects (including the Cross-Center Evaluation work group where the survey originated). Even so, cross-center ties were modest. While the ISC 3 network is a newly funded initiative, a number of the centers' leaders have been collaborating and leading the IS eld for quite some time. We see that with the advanced IS members being highly connected (beyond the fact that they have been cultivating connections longer) and with that expertise, building and facilitating connectivity is all the more important.
No "ideal" number for density exists as a target and it is necessary to balance access to network members with the cost of maintaining relationships. This balance -understanding the interplay between ideal inter and intra density -is an interesting intersection and warrants additional scienti c inquiry. A common measure of inter-intra connectivity is the E-I index which compares ties to members within groups (internal) to those outside of that group (external), though interpreting changes in this measure over time is challenging. Our goal from the activities both planned within and across centers would increase connections in both inter and intra center connections. Future measures could include inter and intra center density to look at how connections change in tandem, instead of inversely to each other. In the interim, year one data helped locate where planned actions, or network interventions (24) should be prioritized and could be leveraged to increase scienti c collaboration or overall "organizational e ciency." Among network activities, practice/policy dissemination collaborations were sparse, with ve-fold fewer ties than planning/conducting research. This may be attributed to stalled dissemination collaborations in the rst year (2020) given community partners of interest pivoted to COVID-19 response. In addition, low collaboration compared to other activities could be due to investigators focusing on intra and inter-network project planning in the rst year of a brand new initiative that more generally precede dissemination to non-science audiences. However, to enhance dissemination to audiences other than researchers, more attention is needed on designing for dissemination (in all stages of the initiative), or the "active process that helps to ensure that public health interventions, often evaluated by researchers, are developed in ways that match well with adopters' needs, assets, and time frames (30,31)." A number of processes could collaboratively be developed and implemented across the ISC 3 including: participatory co-design; context and situation analysis; methods from marketing and business; communications and visual arts; and systems science (31). Such processes could be applied to cross-center projects and also funded pilot work with similar research themes. Processes to determine overlap in research interests can help direct shared, collaborative dissemination. Such information is currently being collected on an annual basis, but establishing systematic communication across centers could be a potential next step for the ISC 3 work groups.
Mentoring the next generation of IS investigators is imperative to "grow the network younger" and to assure that early career members have equal access to collaborative activities and can increase scienti c production overall (32)(33)(34)(35). We posit that early career and underrepresented minority researchers occupying central nodes is essential for a healthy scienti c collaboration network. All centers provide funding for post-doctoral positions. Even if post-doctoral positions are not directly funded by ISC 3 , centers are able to connect early investigators to the larger ISC 3 network, providing a platform to connect with other peers in the IS eld from several other universities. In addition, a work group in Capacity Building formed to increase opportunities and access to resources in the ISC 3 network. In examining isolates in each category, we found that half of the network's trainees were not actively collaborating in cross-network work groups. Providing opportunities for leadership roles to early career and IS trainees could provide for more inclusion in speci c activity networks and integrate them into the broader networks that serve ISC 3 and cancer prevention and control more broadly. Currently, ve work groups exist in the ISC 3 and are speci c places where purposeful engagement of trainees could connect them not only to network activities, but also potential senior mentors outside of their respective centers. Mentoring and access to external experts increased peer-to-peer collaborations (and mentoring) among trainees in an IS training program (33). With longitudinal surveys, we are able to examine any shifts in key connections for others in the network. For example, the NCI Transdisciplinary Research on Energetics and Cancer network demonstrated the utility of social network analysis in measuring changes in dispersion of responsibilities and other network functions through examining brokerage roles (15). Providing leadership roles in work groups, access to experts and pilot funding earmarked for early career investigators or trainees will likely result in shifting network dynamics. In addition, ISC 3 developed a supplemental funding avenue to enhance IS health equity work across the network. Such funding could be a promising mechanism to design features to include less connected, investigators/researchers and also promote cross-center collaborations.
We found that Hispanic or Latino and white members were highly connected across the various network activities. It is important to note that Hispanic or Latino members only make up <3% of the entire network and white members the majority (79%). In general, Black, Hispanic or Latino, Native American, and other groups are under-represented in the ISC 3 network. This points to the need for more efforts to assemble and engage a diverse set of network members. Recently, Leone Sciabolazza and team created a network alteration program that paired previously unconnected researchers within a university through a pilot awards (22). Researchers were introduced and offered a monetary incentive to submit a joint letter of intent for pilot funding. This is one of many opportunities to actively shape the composition and structure of the network.
While these ndings inform strategies to enhance scienti c linkages across the network, important limitations should be noted. First, it is possible that not every network member is positioned or skilled to be involved with every activity that we identi ed and collected information on. Social network surveys are self-report and can introduce some respondent bias, and symmetrizing ties has implications for both respondents who tend to over report and those that under report collaborations. However, our sensitivity analyses demonstrated that extreme cases had little impact on our overall network statistics. It is also possible that we missed people in the network with our center-identi ed roster approach, though with guidance on inclusion criteria, we believe this was likely minimized. Our network analysis included only researchers and given the importance of partnerships and stakeholders to all seven centers, we are not able to examine networks outside the university setting in the current study.
This study illustrates the range of insights offered by a network evaluation for multi-center research initiatives. The analysis highlighted several opportunities to increase participation in cross-center networks and activity-speci c networks (e.g., dissemination to practice/policy audiences). The early evaluation of network participation also provides centers with an opportunity to improve engagement and retention of under-represented groups, including racial and ethnic minorities and trainees.
Future directions include additional social network data collection and comparisons of network activity and growth as part of the outcomes of center-focused initiatives over the ve-year funding period.

Conclusion
We presented baseline scienti c linkages across a robust network of centers working in implementation science in cancer control.
The centers are fairly cohesive and have considerable cross-center collaborations underway. Even so, this snapshot highlights parts of the network where linkages should grow in order for the ISC 3 initiative to meet its objectives including increasing the number of trainees, enhancing practice and policy dissemination, and expanding engagement among members from underrepresented minority groups. Targeted interventions within the network are next steps with plans to use this study as a baseline to measure changes in the network over time. a Density is the ratio of the number of ties to the total number of possible ties in the network; often used to measure the overall connectivity of a network or degree of cohesion among a network of collaborators [0, 1]. b Centralization is used to assess the extent of hierarchy in the network; extent that connections in the network are associated with a select few most central nodes in the network [0, 1]. Degree centralization is based on the number of connections (higher degree centralization=one or more nodes hold most of the connections), whereas betweenness centralization is used to measure the extent to which each network member represents a bridge or gatekeeper to others in the network (based on the number of connections or paths in the network an individual lies between, higher betweenness centralization=one or a few nodes responsible for holding network together). c Transitivity is a measure of clustering [0, 1] with higher transitivity suggests that new ties are more likely to form between nodes that share a common collaborator (e.g. referred by an existing collaborator). Figure 1 ISC3 network of all collaboration activities combined (n=192). Node color represents ISC3 center, node size represents degree centrality scores, and nodes with black borders indicate those reporting "advanced" expertise in implementation science. Square nodes represent those with missing information about IS expertise (n=10).