Article Text
Abstract
Background Interest in the use of social network analysis (SNA) in healthcare research has increased, but there has been little methodological research on how to choose the name generators that are often used to collect primary data on the social connection between individuals for SNA.
Objective We sought to determine a minimum set of name generators sufficient to distinguish the social networks of a target population of physicians active in quality improvement (QI).
Methods We conducted a pilot survey including 8 name generators in a convenience sample of 25 physicians active in QI to characterize their social networks. We used multidimensional scaling to determine what subset of these name generators was needed to distinguish these social networks.
Results We found that some physicians maintain a social network organized around a specific colleague who performed multiple roles while others maintained highly differentiated networks. We found that a set of 5 of the 8 name generators we used was needed to distinguish the networks of these physicians.
Conclusions Beyond methodology for selecting name generators, our findings suggest that QI networks may require 5 or more generators to elicit valid sets of relevant actors and relations in this target population.
Statistics from Altmetric.com
Introduction
There has been a proliferation of articles published in the health sciences literature applying theories and empirical methods of social network analysis to the study of topics ranging from: disease transmission and policy design;1–4 network structure as a determinant of health, health behaviour and well-being;5–8 the diffusion of information and adoption of medical innovation;9–11 and healthcare quality and safety.12–14 This growth of interest in social network analysis reflects increasing recognition that health processes and outcomes are influenced by the web of social connections within which patients and providers are embedded.
Although health research is rich in observational data on social ties and relational processes,1 ,3 ,15 social network surveys remain an indispensable tool for collecting primary data on specific social ties and social processes.5 ,9 ,11 ,12 ,16 At a minimum, a social network survey consists of a name generator, which is a question that asks respondents to enumerate individuals with whom they share a particular type of relationship. Social network surveys also often include name interpreters, which are questions that ask respondents about the attributes (eg, gender, years known) of individuals they named in response to the name generator. Name interpreters provide useful descriptive data, but name generators provide the essential components that define a social network: actors and the relations among them.
The selection of an appropriate name generator and how many generators should be used has received surprisingly little systematic study despite excellent reviews on social network analysis in health-related applications.17 ,18 We report on a methodological pilot study designed to inform the selection of name generators for use in studies of clinician social networks for healthcare quality improvement (QI).
The anatomy of a name generator
A social network is defined as a set of actors and the relations among them.19 A name generator is a survey item that asks respondents to name a certain number of other actors with whom they share a particular type of social relation. For example, in the 1985 General Social Survey (GSS), respondents were asked:
From time to time, most people discuss important matters with other people. Looking back over the last 6 months who are the people with whom you discussed matters important to you? Just tell me their first names or initials.20 ,21
Other examples include asking respondents to name contacts: whom respondents know well;22 who occupy a ‘significant role’ in respondents’ lives over some period of time;5 whom respondents consulted for ‘work-related problems’;16 or with whom respondents interact ‘most frequently’.14
The choice of name generators has implications for the validity and reliability and network data and measures.23 When researchers are interested in studying how social networks affect some social process of interest, a valid network is one that includes all the individuals whose membership affects the process (even if these individuals may occupy different roles), and excludes individuals whose membership has no bearing on the process. Thus, in designing a social network survey for studying a particular process or group of processes (eg, medical care practices), careful consideration should be given to how many and which name generators to use to enumerate relevant actors and relationships among them (eg, trust, advice seeking). Relying on a single role to identify social actors when multiple roles are actually relevant may lead to what Laumann et al24 call the ‘partial system fallacy’, which arises ‘wherever a set of relationships connecting a subset of the actors to which the relations are relevant is analyzed without prior attention to the entire set of actors. The result of such a procedure may be a highly misleading description of network structure’.
There are also practical concerns. The ideal set of name generators for a study covers the domain of interpersonal activities under study without over sampling any one area in the domain. Generators that elicit the same kinds of relations are a burden on respondents and an extravagance when interview time is limited. Methods were developed in Anthropology and Sociology to describe how kinds of relations are distributed in a domain, and those methods were adapted to the task of making meaningful distinctions between kinds of relations.25–,27 For example, the GSS mentioned above is under severe time pressure from competing interests trying to get time on the survey. Each item included has been evaluated for the data it will provide versus the time it will require. Only one name generator was allowed for the social network module. To make effective and efficient use of available time, a name generator was selected by mapping the domain of alternative name generators to identify the one that best represented the whole domain.20 We used the same methods in preparation for a survey of physician networks.
Pilot study
We conducted a pilot study to identify a minimum set of name generators sufficient to distinguish physician networks relevant to QI. Results of this study were intended to inform the design of a subsequent large-scale social network survey to collect data on the structure of interactions among paediatric gastroenterologists and other clinicians within and across sites in a QI network (Improve Care Now (ICN)) in order to guide the future design of QI intervention and dissemination strategies to enhance quality of care for patients with inflammatory bowel disease within the network. Identifying a minimum set of name generators for the subsequent study was felt to be important because we thought that the burden to respondents of using a larger number of name generators would have seriously reduced the practicality and response rates of the planned larger study.
Setting and participants
This study was conducted between March and May 2010. Subjects were 25 physicians at ICN sites who were non-randomly recruited by one of the Principal Investigators and ICN staff. Subjects were sent informational letters and consent forms to indicate their voluntary participation in the research. This study was approved by the Institutional Review Boards of Cincinnati Children's Hospital Medical Center and The University of Chicago.
Social network survey design and administration
The pilot survey was administered as a pen-and-paper survey, and included introductory questions to collect data on basic sociodemographic and professional characteristics on each respondent, including: full name; year of birth; gender; years at practice; board certification specialty area; location of medical school and year of medical school graduation; location of residency; and location of fellowship.
Eight name generators, listed in table 1, were included in the survey, including a general name generator from the GSS. Respondents were asked to enumerate up to five names per generator and to describe each relationship:
-
Does [CONTACT] work at this practice? Yes/No
-
Is [CONTACT] a…Physician/Nurse/Other healthcare provider/Other (please specify)
-
Is your relationship with [CONTACT]…Close/Distant/Something in-between.
Finally, respondents were asked to rate the interpersonal connection between each contact they named: ‘Please describe the nature of the relationship (close, distant, something in-between or don't know) between each of the 5 people you named’. Respondent reports of their perceptions of pair-wise relationships among their contacts is known as enumerating ‘cognitive networks’ and can be a secondary line of defence against missing data from non-response.28
Analysis
Our analysis is guided by the idea that distinctions between kinds of relations are meaningful to the extent that different kinds of relationships reach different people. Two kinds of relations that reach the same people may be confused for one another. Someone whose friends are all outside the workplace sharply demarcates friends from colleagues compared with someone whose friends are also colleagues. We analysed this using a matrix with row dimension equal to the total number of unique individuals named by the 25 respondents, and column dimension equal to the number of kinds of relations distinguished in the interview. In all, 16 kinds of relations are distinguished in our pilot interviews: eight kinds corresponding to the eight name generators (relationship contents) and the following eight kinds of relations distinguished by the name interpreters: colleague; emotionally close; distant; family member; spouse; nurse; administrator; or other. Let i index contacts and j index relational contents/attributes (j = 1,2, …, M; M = 16 in this study). Each cell ij could take on a value of 0 if a row contact i was not named in name generator j; or, cell ij could take on a value of 1 if row contact i was named for name generator (or attribute) j. We computed the conditional probability pjk that a relational content/attribute j and another relational content/attribute k co-occur in a given contact:25
where, njk is the total number of contacts who were cited by respondents for both relational content/attributes j and k, and nk is the total number of contacts who were cited by respondents for relational content/attribute k.
A 16×16 matrix containing these joint probabilities was input into Stata (College Station, Texas) for multidimensional scaling.
Results
Table 1 presents descriptive statistics for the eight name generators. The total number of contacts named by respondents ranged from 9 to 28. On average, respondents named 19 contacts. We used the network data to sort physicians into response categories, then to draw cognitive maps for each category of physicians.
Physicians are sorted into response categories on the basis of figure 1. The left panel plots the total number of contacts (‘colleagues’) who cited a respondent against the maximum times the respondent named any one of their contacts (y-axis, maximum value=8). In social network analysis, the respondent is often called the ego, while the each person they name is called an alter. The large red dots represent physicians who cited one or more colleagues on all eight name generators. The colleague(s) cited across all name generators are ‘anchor contacts’; the respondent's network was ‘anchored’ on the colleague for everything: personal discussion, social activities, treatment options, difficult cases, hospital operations and professional development. Figure 1 reveals few respondents with anchor contacts. Those with anchor contacts were infrequently cited by colleagues. At the other extreme, some physicians were frequently cited by their colleagues and generally maintained specialised relations with other colleagues (hollow dots).
The right panel in figure 1 plots the total number of contacts who cited ego against the average conditional probability between kinds of relations, that is, the relational ambiguity metric introduced in the previous section against the number (y-axis). Physicians with anchor contacts tend to maintain ambiguous relations. For these physicians, each relational content/attribute has a high average conditional probability of co-occurrence with other contents/attributes. Conversely, for physicians maintaining specialised, differentiated networks, any given relational content/attribute tends to have a relatively low average conditional probability of co-occurrence with other contents/attributes. In sum, figure 1 suggests that the physicians we surveyed could be segmented into two groups: (1) those with diffuse and ambiguous relations for whom few name generators would be required and (2) those with specialised relations for whom more name generators would be required to help ensure complete enumeration of relevant contacts for understanding network processes in QI.
Cognitive maps of physicians with anchor contacts
Table 2 presents the matrix of conditional probabilities between the relational contents/attributes for physicians with anchor contacts. Celljk contains the probability that a contact cited for j was also cited for k. As a concrete example, the six physicians with anchor contacts cited 22 people as individuals with whom they discussed QI. Of these 22 contacts, seven were also cited as individuals with whom these six physicians also discussed their professional development. The conditional probability that a contact cited as a QI discussant is also cited as an advisor in professional development is thus 7/22=0.32. Because all 22 contacts cited by physicians with anchor contacts were also colleagues working in the same practice as the citing physician, the conditional probability that a contacted cited for QI is also cited as a colleague is 1.00.
These conditional probabilities were used to produce the multidimensional scaling map in figure 2. Solid squares represent relational contents while hollow squares denote attributes. Figure 2 is a cognitive map in the sense that it depicts contents and attributes in close proximity based on the overall similarity of their patterns of conditional probabilities. Squares for two kinds of relations are depicted in close proximity to the extent that the two kinds of relations often reach the same contacts and are associated similarly with other kinds of relations.29 To facilitate visual interpretation, we deleted lines between contents/attributes with conditional probabilities <0.50. Line thickness is weighted in proportion to the magnitude of conditional probabilities with thicker lines representing higher pjk. Figure 2 shows that physicians with anchor contacts segmented their contacts into two groups: personal relations (clustered to the left) and professional ties (clustered to the right). Once cognitively categorised as a personal contact, all manners of personal social activities tended to be shared with the contact: emotional closeness, social activities and discussing important matters. The cognitive social sphere of personal relations involved family as well as ‘other’ relations that included friends and spiritual advisors. The professional cognitive social sphere tended to be anchored around the role of ‘colleague’, with all professional activities organised around this collegial role: discussion of difficult cases and treatment options, operations and QI, administration, career issues and professional development.
For physicians with anchor contacts, nurses were the most likely targets of both personal and professional relations, and ‘bridged’ the two social spheres. Nurses cited as contacts were the most likely of all other professional contacts to also be cited as people with whom a physician felt emotionally close (PNurse,EClose=0.75, table 2), and were generally cited as contacts with whom the responding physician also discussed operational issues (PNurse,Ops=0.88). Nurses were always cited as colleagues. Physicians with anchor contacts often felt emotional attachment to nurses whom they also cited as colleagues, but the cognitive boundary separating personal and professional relations meant that discussions with their cited nurses were not a source of treatment information or professional development.
A striking feature of figure 2 is the absence of distant relations. Physicians with anchor contacts only cited people for whom they felt some positive level of emotional closeness. No lines connect the ‘Distant’ attribute to any other content/attribute, indicating a very low conditional probability of co-occurrence. In summary, for physicians with anchor contacts, a social network survey need only to include two name generators: the GSS name generator (discuss important personal matters) to obtain personal contacts that are most likely related to a professional discussion and a ‘colleague’ name generator to obtain significant professional contacts.
Cognitive maps of physicians with specialised differentiated contacts: ‘connected physicians’
Table 3 and figure 3 present the conditional probabilities and cognitive map of relational contents/attributes for physicians whose contacts are differentiated across different types of relations (hollow circles in figure 1).
For these ‘connected’ physicians, the GSS name generator would be sufficient to obtain personal relations: all personal relations are strikingly sequestered to the left of the GSS generator in figure 3. Although people named on the SOCIAL (free/leisure time) generator do not overlap much with those enumerated by the GSS generator, it is unclear whether a study of physicians in a specific application such as QI would require social contacts with little professional content. Where the cognitive maps of connected physicians and physicians with anchor contacts differ lies in the types of personal relations that the GSS generator (discuss important personal matters) will elicit. For connected physicians, a contact with whom personal matters are discussed has a 0.57 probability of also being named a colleague, and a 0.46 probability of also being named as someone with whom the responding physician felt emotionally close. By contrast, contacts named by physicians with anchor contacts as someone with whom personal matters are discussed had a much lower probability of (Pjk = 0.40) of also being cited as a colleague, but were more likely to also be someone to whom the responding physician felt close (Pjk = 0.67). These differences parallel differences found within the American population with respect to personal discussion networks: more educated people are more likely to name non-familial contacts on the GSS name generator, and in particular, are more likely to name non-kin colleagues.30
Although a single social name generator may enable enumeration of non-professional social ties for connected physicians, multiple name generators are required to fully enumerate the professional networks of these physicians as they relate to QI activities. As shown in figure 3, professional relations are more differentiated compared with those of physicians with anchor contacts. Connected physicians seem to subdivide their professional relations between contacts associated with career concerns and those associated with clinical practice. Distant contacts were associated with discussing career advancement and professional development. The conditional probability PCareer,Colleague that a contacted cited as a discussant for professional issues would also be a colleague was 0.30 compared with 0.59 for physicians with anchor contacts. The conditional probability PDevel,Colleague that a contact serving as a discussant for professional development would also be a colleague was 0.52 for connected physicians compared with 0.71 for physicians with anchor contacts. Discussions of difficult cases, treatment options, QI and operations were associated with closer collegial relations as well as personal ties with whom respondents engaged in general social activities. Unlike physicians with anchor contacts, connected physicians were more likely to discuss career concerns with individuals who were emotionally and geographically distant: individuals who were contacted for expertise rather than for socio-emotional qualities. This is an important network feature to capture, because distant contacts are more likely than colleagues to be sources of information regarding alternative operational and treatment options. These distant contacts are ‘weak ties’31 that span structural holes.32 Given the satellite positioning of contacts associated with career concerns, a name generator pertaining to career development or professional advancement should be included.
Connected physicians also differ from physicians with anchor contacts with respect to routine clinical operations discussants. In figure 2, physicians with anchor contacts discuss routine clinical operations with administrators, nurses and others with whom they discuss professional development and difficult cases. However, connected physicians discuss routine clinical operations only with nurses. Thus, the ‘operations’ name generator will help ensure that nurses and targets for operational discussions are enumerated in the networks of connected physicians.
In figure 3, the bold lines from Colleague to DIFF (talk to about difficult cases) and TREAT (talk to about treatments) suggest that QI does not occupy a central position in the cognitive maps of connected physicians. In table 3, half the contacts cited as QI discussants were colleagues in the same practice (PQI,Colleague = 0.52), but QI discussants represented only a fifth of all colleagues. The seven connected physicians cited 25 contacts with whom they discussed QI; the six physicians with anchor contacts cited 22 contacts. Connected physicians were more likely to cite contacts outside their own practices as QI discussants. Thus, in a network study of QI, a specific name generator to elicit names of individuals with whom QI is discussed is important to include because connected physicians seem to have greater access to extramural colleagues in this content area.
Conclusions: lessons learned/implications for research
Methods previously used to select name generators in national probability surveys were applied to effectively and efficiently survey physician networks. Our pilot study shows that reliance on a single, general name generator to collect social network data for applications related to QI may be inadequate for enumerating valid QI networks. The results of our study suggest that multiple name generators may be necessary to ensure complete enumeration of all ties that may be potentially relevant to understanding network processes in QI. Specifically, this pilot suggests we should include five generators in the future survey: the GSS, colleague, career/professional development, operations generator and QI generators. The generalisability of our findings to other QI contexts is unclear. Future replications of this brief pilot study would be helpful to determine how robust or generalisable our findings are across different study populations. In addition, even in the specific context, it should be noted that name generators that requested more or less than five contacts might well have led to a different set of name generators being identified. Since increasing the number of contacts requested increases respondent burden, there will often be a trade-off between the number of respondents requested and the number of name generators that can be studied. Exploring the robustness of our results to alternate numbers of respondents and/or name generators are also important areas for future work in the methods for selecting name generators.
Acknowledgments
The authors gratefully acknowledge funding for this research from the National Institute of Diabetes and Digestive and Kidney Diseases (‘Open Source Science: Transforming Chronic Illness Care’, 5R01DK085719-02, Drs Margolis and Seid, Principal Investigators).
References
Footnotes
-
Contributors All authors made substantial contributions to conception and design, acquisition of data, or analysis and interpretation of data. All authors contributed to drafting the article or revising it critically for important intellectual content. All authors provided final approval of the version to be published.
-
Funding We received funding for this research from the National Institute of Diabetes and Digestive and Kidney Diseases (‘Open Source Science: Transforming Chronic Illness Care’, 5R01DK085719-02, Drs Margolis and Seid, Principal Investigators).
-
Competing interests None.
-
Ethics approval Ethics approval provided by the University of Chicago Institutional Review Board.
-
Provenance and peer review Not commissioned; externally peer reviewed.
-
Data sharing statement Due to confidentiality considerations, compliance with our Institutional Review Board precludes any data sharing.