How to do (or not to do) . . . a social network analysis in health systems research

The main challenges in international health are to scale up effective health interventions in low- and middle-income countries in order to reach a higher proportion of the population. This can be achieved through better insight into how health systems are structured. Social network analysis can provide an appropriate and innovative paradigm for the health systems researcher, allow new analyses of the structure of health systems, and facilitate understanding of the role of stakeholders within a health system. The social network analysis methodology adapted to health systems research and described in detail by the authors comprises three main stages: (i) describing the set of actors and members of the network; (ii) characterizing the relationships between actors; and (iii) analysing the structure of the systems. Evidence generated through social network analysis could help policy makers to understand how health systems react over time and to better adjust health programmes and innovations to the capacities of health systems in low- and middle-income countries to achieve universal coverage.


Introduction
Since the launch of the Millennium Development Goals (MDGs) in 2000 (United Nations 2000), health systems have become a priority focus for researchers, managers and decision makers, and efforts have been increasingly invested to improve the capacity of local health systems with the aim of achieving the MDGs by 2015 (Murray and Frenk 2000;Sachs et al. 2004;Reich et al. 2008). The message from the international experts who came together as part of the 2008 G8 Summit held in Japan was very clear: health system strengthening and disease-specific strategies were no longer seen as competing approaches but were complementary interventions (Fukuda 2008;G8 Health Expert Group 2008).
However, today, researchers, managers and decision makers face two major challenges with regards to health systems: describing with accuracy the subject of their study and understanding how the structure of health systems can influence health outcomes (Figueras et al. 2008;Jensen 2008). What is a health system? And how do dynamic and diverse health systems influence the delivery of health services and the production of health outcomes? In health systems research, social networks have implicitly been at the heart of the definition of health systems (Merrill et al. 2008;Wholey et al. 2009). According to the World Health Organization (2000), a health system is defined as all the organizations, people and actions whose primary intent is to promote, restore or maintain health. Kohn et al. (2000) made the link between social networks and health systems even more explicit in their own definition of the health system. They saw a health system as a network of actors who aim to provide health care: 'In health care, a system can be an integrated delivery system, a centrally owned multihospital system, or a virtual system comprised of many different partners over a wide geographical area' (Kohn et al. 2000, p. 52).
In a globalized world, health systems in low-and middleincome countries (LMIC) have become multi-scale and dynamic entities. In the context of this paper, scales are defined as the spatial, jurisdictional and administrative dimensions used to study the structure of the health system (Ostrom et al. 1999;Gibson et al. 2000;Cash et al. 2006). For example, regional health authorities have the responsibility to monitor the health situation of the whole region but need to collect data from different spatial and administrative scales of the regional health system (villages, sub-districts and districts). In a health system, scales can be defined by the catchment areas of every health organization involved in the delivery and management of health care. The interconnections between actors from different spatial, administrative or jurisdictional scales are the result of the multiplicity and diversity of actors intervening in the health sector and the close interactions between traditional and modern medicine, between formal and informal sectors, between international, national, regional and local actors, and between public and private sectors (Bloom 2001;Mackintosh and Koivusalo 2005). In addition to this pluralism, the development of information technologies has accelerated the interactions between global health systems and local health systems, and increased the dynamics of health systems (Morgan 2005). An issue that emerges from this new situation is how the nature and structure of health systems might be captured in such a complex environment. In the race to reach the MDGs, it has been recognized that the capacities of most health systems constituted key obstacles to generate effective and equitable health outcomes amongst the populations of LMIC (Waage et al. 2010).
In the present paper, the authors will demonstrate that seeing health systems as social networks can help to define the boundaries of health systems and understand the correlation between health systems' performance and network parameters in LMIC. Better understanding of health systems will improve the effectiveness and coverage of health programmes in developing countries (World Health Organization 2010a).

Social networks in health care
Social network analysis (SNA) is defined as a distinctive set of methods used for mapping, measuring and analysing the social relationships between people, groups and organizations (Scott 1999;Borgatti et al. 2009). Using mathematical algorithms (Marsden 1990) and software (e.g. UCINET) (Borgatti et al. 2002), researchers have analysed how patterns of relationships between actors within a system can facilitate or constrain the individual decisions and actions of actors, as well as system functions and adaptive capacities (Wasserman and Faust 1994). In a graphic representation of social networks (Figure 1), SNA illustrates an actor (e.g. an individual, a family, a community or an organization) by a node and the relationships between actors by ties (Marsden 1990;Degenne and Forsé 1999;Borgatti and Cross 2003;Batley and Larbi 2004;Islam 2007). Relationships between actors can be as diverse as friendship, trust or knowledge transmission (Folke et al. 2002).
Beyond being a method, SNA is also viewed as a paradigm that has its own international society (the International Network for Social Network Analysis) and its own scientific journal (Social Networks). SNA provides an avenue for analysing and comparing formal and informal information flows in a system. SNA recognizes the complexity and dynamics of networks and their influence on behaviour and decisions (Borgatti et al. 2009). SNA has proved that it can be used to help understand the nature of relations between actors within a system and how these relationships influence the structure of a system (Webb and Bodin 2008;Borgatti et al. 2009). Although SNA and health care have long been interconnected, SNA has not been applied yet to health systems research in LMIC, which remains a nascent field of investigation.
Social network theories have a long history in health care. Social network theories were born in public health in 1934 when, after an epidemic in a New York school, Moreno tried to understand why the epidemic spread so quickly amongst the pupils (Moreno 1934). Moreno (1934) was also the first to represent graphically the relationships between pupils and their social position with each other. In order to model systems, social networks theoreticians applied mathematical and graphical techniques to illustrate and understand the complexity of human and organizational relationships.
The role of networks has become crucial in health care during the 21st century with the emergence of informational and technological innovations, and with the recognition from health managers that hospitals were no longer the unique place where health care was delivered (Greenhalgh 2008). Health care providers have acknowledged the role of other actors-medical and non-medical, private and public-and the positive impact of multi-scale and multi-disciplinary network-based initiatives involving medical staff working in hospitals, health staff posted in primary care health facilities or community-based workers (Atkinson 2002;Bloom and Standing 2008).
Applying SNA in health systems research encounters a number of challenges, such as capturing the dynamics of systems, the limits of a social network and the effects of multi-scale events that affect several spatial scales of the health system (e.g. the increase of the price of oil has an impact on the delivery of health care services at regional, district and community levels). Therefore, innovative approaches were SOCIAL NETWORK ANALYSIS IN HEALTH SYSTEMS RESEARCH introduced combining social network theories and other approaches that could potentially generate new knowledge when applied to health systems (Cumming et al. 2010). At the individual level, social scientists showed that social networks determined the level of co-operation between individuals: individuals tend to collaborate more easily with their direct neighbours. SNA researchers also showed that, although individuals are connected with a limited number of people, people in the world are all indirectly connected by a number of ties that on average does not exceed 'six degrees' (Watts and Strogatz 1998). This high degree of connectivity between individuals and organizations has implications for the level of interdependence and embeddedness between networks. Individuals connected through a social network tend to have similar beliefs and values (McGuire 2000;Kiesler and Cummings 2002;Uzzi and Gillespie 2002). Scholars found that there was a relationship between the structure of networks, the type of links between actors (i.e. bonding between actors of the system or bridging links with other systems) and the resilience of social-ecological systems (Folke et al. 2005). At the network level, SNA has also been used to analyse the patterns of diffusion of innovations and in particular how the structure of a network or system determined the degree of adoption of innovations. This was applied in various fields such as farming or business (Rogers 1995).
Literature on social networks is now relatively vast. SNA has had concrete applications in many fields including health behaviour, health prevention, organizational management or group behaviour (Valente 2010). How SNA can serve the interests of health systems researchers by providing concrete measures and tools to define health systems is described in the following sections.

The main characteristics of social networks and health systems
Health systems research aims to understand health governance in a context characterized by a multitude of diverse actors (World Health Organization 2009). Governance is one of the six functions of health systems along with service delivery, financing, human resources, technology and health information systems (World Health Organization 2010b). Lebel et al. (2006) proposed a conceptual framework to describe the six main characteristics of the 'good' governance of social-ecological systems. Of these six characteristics, three can be applied to the governance of health systems: (i) capacity to engage effectively with and handle multiple-and cross-scale dynamics; (ii) capacity to anticipate and cope with uncertainties and surprises; and finally (iii) capacity to combine and integrate different forms of knowledge. To illustrate the utilization of network tools to health systems, the authors will show how these three system properties can be analysed by using five different network properties: two properties related to the structure of the network and three properties related to the position of actors.
First, in terms of general structure of a network, two properties are particularly used in SNA: cohesion and shape (Borgatti et al. 2009). Cohesion describes the number of connections within a network and includes sub-properties such as density and fragmentation. More dense networks have a higher number of connections between actors. Shape relates to the overall distribution of ties and distinguishes the core actors from the peripheral ones (Borgatti et al. 1990). The core actors are highly connected with each other while the peripheral actors have loose links.
A second application of SNA in health systems research is in the analysis of the role and position of specific actors. Health systems research focuses on the role of actors within a health system (World Health Organization 2009) and in the diffusion of knowledge and innovations, such as in the management of epidemics (Rogers 1995;Latkin et al. 2003;Helleringer and Kohler 2005). However, work in health systems research is still at an early stage and the available tools such as stakeholder analysis (Glassman et al. 1999;Brugha and Varvasovszky 2000) provide limited analysis of the role of actors in a health system. SNA could be an appropriate analysis tool to generate an actor-level analysis of health systems (Borgatti and Foster 2003). SNA can also be a valuable tool to uncover the most influential players in a system (Valente and Pumpuang 2007;Riggan and Supovitz 2008).
One finding from SNA is that the position of an actor in a network determines their capacity to access and diffuse knowledge and information or, in other words, control the flow of information (Borgatti et al. 2009). SNA provides tools to identify a knowledge broker, i.e. individuals who create links between users and researchers (Thompson et al. 2006).
The brokers in a health system will help co-ordinate actors in times of crises or shocks and build bridges between different groups of the system (Burt 2003;Newman and Dale 2005). Other actors essential to the diffusion of innovations, such as opinion leaders, champions or change agents, can be identified through the number of links they have with their peers or non-peer actors at different levels of the health system (Berner et al. 2003). For example, opinion leaders, i.e. people who can influence other people 0 s views (Rogers and Cartano 1962), have the highest numbers of ties within a network. Identifying opinion leaders and building a programme through these key people can help to diffuse innovations in a network, e.g. utilization of medical guidelines (Lomas et al. 1991) or HIV risk-reduction practices (Sikkema et al. 2000). Centrality, reachability and betweenness are the most well-known node-related properties (Freeman 1977). The definitions of these quantitative measures are presented in Table 1.

How to design a social network analysis
The methodology described below is one approach to SNA, but other study designs could be used and explored by interested researchers. In the following sections, particular attention is paid to adapting SNA design to health and more specifically to health systems research in LMIC. The methodology presented in this paper was elaborated and tested by the authors in a study conducted between 2008 and 2010 in the Brong Ahafo region of Ghana. That study aimed to analyse the influence of district hospital directors' social networks on their capacity to make management decisions.
The SNA methodology developed by the authors consists of three main stages: (i) describing the set of actors and members of the network; (ii) characterizing the relationships between actors; and (iii) analysing the structure of the systems (see Box 1).

Describing the list of actors and members of the network
The first stage of SNA consists of describing the actors and members of the network. Actors are defined as persons, informal groups of people or formal organizations who may

Betweenness
Betweenness is a measure that indicates how much a node is located in the path between other actors or how much a node connects other nodes with each other (Freeman 1977).

Centrality
The degree of centrality represents the number of ties a node has (Freeman 1979). If a node has many ties compared with actors, this indicates that this node has a central position in the network. Centrality can also characterize the shape of a whole network.

Density
Density is defined as the number of existing ties divided by the number of possible ties.

Distance
Distance measures the number of ties that separate two actors. If two nodes are directly connected, the distance is one. If these two nodes are separated by one node, the distance is two.

Reachability
Reachability defines the degree by which a node can be reached by other nodes. If a certain number are unreachable by some actors, it means that the network is fragmented. Reachability corresponds to the number of steps maximally needed to reach from one node to any other node in the network.
influence a project's outcomes and the system's resilience both through their interactions, and through individual or collective actions (Freeman 1979;Grimble and Wellard 1997;Brugha and Varvasovszky 2000). The pluralistic nature of health systems means that the diversity of actors involved in a health system is broad and that the boundaries of the system can remain blurred (Bloom et al. 2007). The actors involved in a system can be identified by combining two different but complementary methods proposed by Grimble and Chan (1995): (i) the list of stakeholders involved in a system can be pre-defined by the researcher based on a detailed review of project proposals and documents; (ii) this list of actors is complemented by information collected through interviews with key respondents.
To increase the validity of findings and reduce the incidence of recall bias (e.g. making sure that any actor part of the system will not be omitted), a third step was added by the authors: every interviewee is asked to identify additional actors on the basis of their answers to the following questions adapted from Salam and Noguchi (2006): (a) who gained or lost during the health intervention? (b) Who is expected to gain or lose as a result of the health intervention's success? A final set of actors that are involved in the project is established. These actors become the key informants of the study.

Characterizing the relationships between actors
Relationships between actors can be of different kinds. They depend on various social factors such as trust, conflicts or knowledge sharing (Wasserman and Faust 1994;Folke et al. 2005;Manring 2007). However, all these factors rely on a key process: the circulation of information between and within social networks (Bodin et al. 2006;Manring 2007). Studying information flow mechanisms between actors and within networks can help to understand the social processes influencing health system dynamics and reactions.
The second stage of SNA consists of identifying the existence of flows of information between actors or, in other words, the demand (receiving information) and supply (providing information) of information between individuals. This information can be collected through interviews. The people interviewed are the actors identified during the first stage of the process. A robust method for generating self-reported ties is to use recall lists (Marsden 1990): a list of all organizations in the field with adjoining empty columns in which respondents can mark their different relations to others (Diani 2003). An alternative is to use a paper card for every actor. The pieces of paper are displayed in front of each interviewee.
The interviewee is asked about the demands for information: Do you receive information from this actor? If the interviewee answers 'Yes', then the investigator can ask additional questions to collect more qualitative information about the type of information received. For example, what kind of information do you receive? What is the frequency of your contacts? How do you receive it: by phone, through visits, letters. . .?
The same questions are systematically asked about every actor identified. Once this is completed, the investigator starts again at the beginning of the pile of cards (or the table) and asks about the supply of information: do you provide information to this actor? If the answer from the interviewee is 'Yes', the investigator can ask more questions about the type of information provided and the way the information is circulated.
Data collected are recorded in the information flow matrix elaborated by Brinkerhoff (2004) (see example in Table 2): one matrix for demand of information and a second one for supply of information. Each respondent thus generates a row of 1's and 0's for each of the two network relations (demand and supply of information): '1' symbolizing the existence of demand/supply of information, and '0' signifying no information flow between the two actors.
The third stage in the proposed method consists in analysing the ties between actors and the structure of the network.
Analysing the structure of systems: the use of software packages The properties of the networks are analysed using the algorithms available through specialized software packages Box 1 The three main stages of the social network analysis applied to health systems research Stage 1 Defining the list of actors and members of the network (i) Step 1: List all stakeholders involved in a system based on a detailed review of project proposals and documents; (ii) Step 2: Complement the list of actors with information collected through interviews with key respondents.
Stage 2 Defining the relationships between actors (i) Step 1: Display the list of actors in a table; (ii) Step 2: Interview key informants to identify the relationship between actors; (iii) Step 3: Indicate in the table the existence or absence of a relationship between actors. In each square of the table, a '0' is written when there is no supply of and no demand for information between two actors. The square is filled with '1' when there is a flow of information between the two actors (either a demand or supply of information).
Stage 3 Analysing the structure of the system: measuring five key network properties with the help of the UCINET software: (i) Betweenness (ii) Centrality (iii) Density (iv) Distance (v) Reachability such as UCINET 6, designed by academics for research purposes (Borgatti et al. 1999). Through this method, two matrices are generated: one matrix for demand of information and one matrix for supply of information. The two matrices are then combined to generate a single matrix (Marsden 1990). The final matrix is the result of the addition of the two links. In summary, the new link is N ¼ A þ B, A and B being the value of the link in the demand matrix and the supply matrix. The new network is transformed into a symmetrical and dichotomized network (i.e. without direction of links and no strength, just zeros and ones). The new matrix of the system is inserted into the software UCINET (Borgatti et al. 2002) that helps analyse the properties of the network.
Calculations are run by the software. The network measure calculated is then analysed to understand how health systems are governed. For example, in order to be able to find multi-scale solutions to multi-scale problems, the actors of a network need to be able to get access to information from various types of actors and not only from their close colleagues and neighbours. This means getting access to stakeholders who have access to different sources of information and different types of power (Oh et al. 2004). Access to various sources of information requires a high level of reachability and short distances between actors. Table 3 describes the links between network measures and the systems' properties.
As a final result, health systems are then represented by graphs showing the nature of actors involved and the relationships between these actors. The analysis provides valuable information to decision makers and managers on what key influent actors were excluded from the systems and what new relationships should be encouraged to facilitate collaboration between key players. The analysis of the structure of health systems provides information on the properties of the system (e.g. density, centrality). Describing the strength and weaknesses of the structure of health systems can help decision makers predict how an innovation can be diffused within the system and what is the best strategy to adopt to circulate information.

Conclusion
Health systems are seen as a combination of various systems embedded within each other, such as public and private systems, local, regional and global systems or social and organizational systems (Snijders and Doreian 2010). An event at the level of one sub-system can have an impact on another sub-system and influence the behaviour of network actors. The complexity and embeddedness of health systems create very similar challenges for analysis to the ones generated by social networks (Laumann et al. 1983). Today, the main priorities in international health are to scale up effective health interventions in LMIC in order to reach a higher proportion of the population. The present paper described the origin of SNA, the added value it can represent in health systems research, health Table 3 Features identified as important for the adaptive management of natural resources and the ways in which they are linked to social network structure

Characteristics of health systems
Corresponding social network variables Capacity to engage effectively with and handle multiple scales Reachability: A measure that describes the capacity to reach many actors to get access to or circulate information (Oh et al. 2004).
Distance: The shorter the distance between actors, the faster the diffusion of information. Distance between actors is calculated by the number of ties separating two actors.
Capacity to anticipate and cope with uncertainties and surprises Centrality: A network with a central structure has more capacities to co-ordinate actors and provide a rapid response (Leavitt 1951;Fujimoto et al. 2009).
Betweenness: Rapid response can only be obtained when actors are quickly informed of events or shocks. This requires close links between actors to quickly diffuse information (Granovetter 1973;Reagans and McEvily 2003).

Capacity to combine and integrate different forms of knowledge
Reachability: The diversity of knowledge can be achieved through relationships with actors that belong to other spheres or other sub-networks (Steel and Weber 2001).
Density: Actors in a dense network have more difficulties accessing diverse forms of knowledge as most actors have very similar backgrounds and values (Granovetter 1973;Frank and Yasumoto 1998;. Table 2 Example of information flow matrix showing the supply of information between actors. The actors listed in column 1 supply information to actors listed in row 1. '0' represents 'absence of supply of information' and '1' represents 'existence of supply of information'.
Supply of information between actors listed in column 1 with actors listed in row 1 User Regional Directorate

Regional doctor
Hospital manager

Community-based organization
User 0 0 1 1 Regional Directorate 1 1 0 Regional doctor 1 0 Hospital manager 1 Community-based organization SOCIAL NETWORK ANALYSIS IN HEALTH SYSTEMS RESEARCH service management and health policy, and how it can be used to analyse the relationships between actors and the social position of actors and their degree of influence. Evidence generated through SNA could help policy makers understand how health systems react over time and how ties between actors can influence the diffusion of innovations. However, additional issues need to be addressed by researchers when studying health systems in LMIC: (i) the dynamics of systems and networks (actors and relationships between actors are in constant evolution); (ii) the factors that determine the structure of health systems and are correlated to contextual factors; (iii) the relationships between the structure of a health system and the structure of other systems, in which the former is embedded in. Progress can be made in health systems research with the help of SNA and in-depth insight can be brought to make better sense of how health systems react to shocks.

Funding
The study was funded by the Swiss Red Cross.