Exploring Sociocentric and Egocentric Approaches for Social Network Analysis

Transcription

1 Exploring Sociocentric and Egocentric Approaches for Social Network Analysis Kenneth K S Chung School of Information Technologies University of Sydney NSW 2006, Australia ken@it.usyd.edu.au Liaquat Hossain School of Information Technologies University of Sydney NSW 2006, Australia lhossain@it.usyd.edu.au Joseph Davis School of Information Technologies University of Sydney NSW 2006, Australia jdavis@it.usyd.edu.au ABSTRACT In this paper, we introduce social network analysis for investigating the effect of network position and Information and Communication Technology (ICT) use on the performance of general practitioners (GPs) residing in rural Australia. Here, we highlight the data collection procedure, its benefits and limitations and standard measures of social network data. We first suggest that collection and analysis of relational and attribute data offers richer insights with regards to the investigation of structural effects of network position on performance of GPs. Second, drawing from theoretical and methodological strengths of previous studies conducted in health care, we describe the process for collecting relational data from GPs in isolated and rural settings. We primarily discuss two specific types of social network data collection approaches--(i) whole network and (ii) ego-centric network approach; and highlight its opportunities and challenges. Third, we discuss the problem of network sampling in the context of GPs in rural areas. Finally, challenges associated with analysing social network data are presented. Keywords Social network, egocentric, sociocentric, rural general practitioners, performance 1. INTRODUCTION Social network studies have gained significant recognition in terms of both theory and method in recent years [18]. Theoretical studies on social networks have greatly impacted various domains such as social capital, knowledge management, network organisations and so on [7]. Based on the theoretical constructs of sociology, mathematical foundations of graph theory and recent developments in computer hardware and software, social network analysis (SNA) offers a unique methodology for visualizing and investigating social structures and relations [39]. While a general social survey usually allows for studying individuals properties as the prime context for explaining outcome, SNA incorporates the social context to explain individual or group outcomes. The relationships between the actors hence become the focus of study and the properties of the actors themselves remain secondary. This paper focuses on the data collection procedure, its benefits and limitations and standard measures of social network data within the context of the knowledge-intensive work of GPs in rural and isolated settings. The development of the field of social networks was brought about in the 1930's by several groups working independently in different traditional fields [18]. In the 1930s, a systematic approach to theory and research began to emerge. Georg Simmel constructed a theory that explained the causes of social phenomena and contributed towards formal sociology, which was the predecessor to SNA. In 1934, Jacob Moreno became the first to operationalise a social network by creating a system for representing a social network as a combination of nodes and links. Later, Dorwin Cartwright and Frank Harary built upon Moreno s work, by applying the concepts of graph theory to the sociogram. They were able to incorporate much more complexity in the patterns of social relations, by adding direction to the lines and to show relationships as positive or negative. By the late 1930s, two separate traditions for SNA had developed [38]. The first was the work by a group at Harvard University on ways to find subgroups of people in larger groups. The sociocentric approach developed from this tradition. It involves the quantification of relationships between people within a defined group. The focus is on measuring the structural patterns of those interactions and how those patterns explain outcomes. The second tradition originated from group of anthropologists at the University of Manchester which paved the way for community studies and gave rise to the egocentric approach. They studied the networks of relations surrounding individuals rather than focusing on the whole society. Therefore, with its focus on individuals, it was concerned with making generalizations about the features of personal networks. 2. SOCIAL NETWORK AND ANALYSIS A social network is basically a set of actors and relations that hold these actors together. Actors can be individuals or aggregate units such as departments, organizations, or families. Actors form social networks by exchanging one or many resources with each other. Such resources can be information, goods, services, social support or financial support. These kinds of resource exchanges are considered a social network relation, where individuals who maintain the relation are said to maintain a tie [14]. The strength of their tie may range from weak to strong, which depends on the number and types of resources they exchange, the frequency of exchanges and the intimacy of the exchanges [33]. Further, social ties consist of multiple relations (as in the case of doctors who have a doctor-patient relationship as well as a friendship relationship) and therefore are called multiplex ties [26]. Social networks can be categorised into one of the two types of networks. One-mode networks consist of a single set of actors. They differ from two-mode networks in that two-mode networks consist of two sets of actors or one set of actors and one set of events [25]. The simplest form of a social network consists of actors and/or events and their connections to each other. The network sociogram presented in Figure 1 depicts the collaboration 1

2 network of social network authors who have co-authored journals together. The nodes in the network represent the authors (coded in numbers) while the relations (links) show the co-author relationships between the authors. The network is one-mode because it is a co-authorship network only. Figure 1: An example of a Social Network [37] The relation in this case is undirected, which means that if author 30 has co-authored with author 35, then the converse is also true. If the relations indicated friendship, then a directed line from actor A to actor B is not necessarily the same as a directed line from actor B to actor A. In the sociogram, author 49 is the most prolific author because he has published the most articles. A simple count of the lines (degree count) indicates that he has co-authored with nine other authors. The component created by the subgraph of author 23, 32, 43, 6, 57 and 25 indicates a clique because the structure shows that each actor within the clique is directly connected to any other node in the subgraph. An interpretation of the clique structure is that the authors were probably interested in the same aspect of the subject of social network analysis which resulted in co-author collaborations and publications. The use of SNA methods depends on the availability of relational rather than attribute data [38]. It further allows the mapping of relationships between people. This can be used to identify knowledge flows such as who do people seek information and knowledge from? Who do they share their information and knowledge with? An organisation chart, shows the formal relationships such as, who works where and who reports to whom, however a social network analysis chart shows the informal relationships such as who knows whom and who shares information and knowledge with whom [13]. As such, SNA makes it possible to visualize and understand the relationships that may facilitate or impede knowledge creation and sharing. Since SNA allows these relationships to be visible, it is sometimes referred to as an organisational x-ray which shows the real networks that operate underneath the surface organisational structure [2]. 3. CONTEXT OF THE STUDY The collection of data from a social network is similar to the collection of data from a sample of individuals in any general social survey (GSS). However, one key difference is that in designing a social network survey, relational data is collected along with attribute data. Burt [10] makes a strong argument as to why relational data should be collected in the administration of a general social survey. First, any national GSS is sociology s premiere database of national survey data that can be used for theoretically informed empirical research. This is the same as with the Australian Bureau of Statistics. Second, the costs are 5 to 11 minutes of interview time in addition to data processing time. Third, and more importantly, the inclusion of the relational data allows for describing and understanding significant aspects of an individual s interpersonal and social environment, social participation, and exposure to normative pressures. Fourth is the analysis of both attribute and relational data (eg. ethnic background and number of social ties in a workforce), which can offer richer insights to explain social outcomes. For example, in a study of directors of nursing and clinical directors by West et al [44], the socio-demographic characteristics were correlated with the relational characteristics to understand differences in their interaction patterns. One of the outcomes of the study was that the clinical directors who were graduates and had undergone higher education were more likely to be associated with professional associations as compared to directors of nursing. The combination of both relational and attribute data therefore offers a very useful way to understand individual outcomes in a given social setting. The domain of this study is the rural General Practice of NSW, Australia. In this study, we explore the effect of social network position and ICT use on performance of rural GPs. GPs working in rural areas are geographically isolated from more populated practices. Therefore, nature of their work, isolation from the urban community, and the numerous problems that plague their practice makes this study potentially important. The collection of primary social network data from rural GPs entail problems related to sampling and mode of administration (face to face interviews, mail outs, telephone interview, etc). Problems such as decreasing GP performance as they age, lack of association with professional peers, keeping up to date with modern technology, isolation from community provide the motivation for an understanding of the interplay between social networks, ICT use and performance of rural GPs. In this paper, we focus on the phase of collecting social network data from a random sample of GPs from rural NSW divisions of General Practice in rural NSW. The issues of sampling (general and network) are discussed below. 4. NETWORK DATA COLLECTION There are two main approaches to social network data collection whole network and egocentric network approach. Below is a brief description of these approaches. 4.1 Whole or Sociocentric Network Approach The focus of a whole network analysis is on measuring the structural patterns of those interactions and how those patterns explain outcomes, like the concentration of power or other resources, within the group. The underlying assumption is that members of a group interact more than would a randomly selected group of similar size. Sociocentric network analysts are interested in identifying structural patterns in cases that can be generalised [42] [21]. 2

3 In a whole network study, the actors of the network are usually known or easily determined. This is because a sociocentric network study usually focuses on closed networks implying that the boundaries of a whole network are a priori defined. In many cases, this approach remains the gold standard because of its ability to gather data for the entire network. The network represents the saturation sample of interests and the analysis allows for the results to be generalised to the population. Data collection using a whole network approach usually involves listing the names of the actors in the form of an adjacency matrix. When respondents are administered the network survey consisting of the roster of names, they usually check off the names of people whom they know depending on the name generator question asked. For example, in a general practice that consists of 15 workers (n=15) including the GP and the administration staff, a whole network study may be conducted in order to understand the communication network of the practice. A roster of the names of all the workers (excluding the respondent s) in the practice will be presented to each of the worker. A simple name generation question such as In the past two months, who have you communicated with more than twice in a week within the practice in order to carry out your daily task? Obviously, a definition of what constitutes a daily task will need to be provided to the respondent. An example of a well-applied whole network approach was used by Hirdes and Scott [27] who undertook the study in a chronic care hospital where random samples of patients, family members and employees were asked to report on relationships with people from each of the constituent groups. The setting in which the whole network study was conducted is interesting because the patients included those suffering from dementia. Given that some patients suffered a high degree of dementia, it was impossible to elicit any form of attribute or relational data from them. The advantage with the whole network approach towards data collection in this sensitive setting is that it allowed investigators to include them in their study. There are challenges for conducting network data collection using a whole network approach in the context of rural GPs. In the study by Hirdes and Scott [27], the boundaries were defined in terms of the patients, families and employees who stayed, visited and worked in the hospital. To conduct a whole network study of the GPs in rural NSW would mean that all names of the GPs would need to be known, which would generate a huge list of names (roster) for recall by the respondent. However, previous studies suggest that scrutinising through long lists of names and identifying the multiple types of ties with each person on the roster causes fatigue and recall problems [5]. Given these difficulties, we explore the egocentric network approach as an alternative strategy for data collection. 4.2 Egocentric Network Approach In the social network parlance, the person we are interested in is referred to as the ego and the people referred to by the ego as his affiliate, advisor, friend, or relative, are known as alters. Coleman, Katz et al [11], in an attempt to understand the underlying social processes of doctors that affect their rate of adoption of a drug called gammanym used egocentric SNA to understand the social structure that linked the doctors together. They explained the drug adoption process empirically by using both relational and attribute data. They included all local doctors in the sample in whose specialties gammanym was of potential significance. This ensured that alters named by the doctors were also included in the sample as a result of which social ties amongst the doctors could be determined. 125 general practitioners, internists, and pediatricians were interviewed and this constituted 85 per cent of the doctors practicing in the four cities of interest. Regarding relational data, three questions were asked to each doctor: (1) to whom did he turn for advice and information? (2) with whom did he most often discuss his cases in the course of an ordinary week? (3) who were the friends, among his colleagues, whom he saw most often socially? While the first two questions are name generator and name interpreter items respectively, the third is a significant one because it elicits firstly, the ties and relations between the doctor s contacts themselves, and secondly, the strength of the relationship between the doctor interviewed and the doctor s close social contacts. With this information available, the construction of the social network of the doctors becomes complete. Attribute data of doctors such as age, number of medical journals subscribed to, attachment to medical institutions outside of the community and certain attitudes of the doctors were also collected. The attribute data (about doctors attitudes) collected allowed for the categorisation of whether a doctor was more patient-oriented or more profession-oriented. It was found that doctors who were generally more profession-oriented were quicker to prescribe the drug as compared to those who were patient-oriented. More importantly, the relational data suggested that doctors who were generally more integrated with their peers were faster in the adoption of the new drug as compared to those who were more isolated. A later study explored the professional affiliation and occupational status of 50 Clinical Directors of Medicine and 50 Directors of Nursing in the UK and claimed that these individual attributes affected the characteristics of their social networks to a certain extent [44]. The two professional groups were chosen because of their probability to have well developed networks and because they were the two most important groups in the medical community in terms of number and power. The authors stratified their sample according to the Binley s Directory of NHS Management and randomly selected the Directors for their study. Similar to Coleman, Katz et al s [11] study, they asked the following name generator question: From time to time, people discuss important professional matters with other people. In the last twelve months, who are the people with whom you have discussed important professional matters? Important professional matters included both clinical and managerial issues and were explained to the respondents accordingly. From each of the respondents, five alters were elicited, and socio-demographic data about alters, including the nature of the relationship between the alters were also collected. The research question that motivated the paper was whether network features (density, centralisation and centrality) were related to structural location in the organisational hierarchy or whether they were simply a function of individual characteristics. Some of the hypotheses that stemmed from the research questions were: 3

4 1. Directors of Nursing were more likely than clinical directors of medicine to name alters who are junior to them; 2. The networks of nursing will be lower in density than clinical directors of medicine; and, 3. Networks of directors of nursing will be more centralizes (as measured by group degree centralisation) than those of clinical directors. 4. Directors of Nursing will have higher actor information centrality scores than clinical directors of medicine. West et al [44] utilised descriptive statistical measures of crossclassifications of the different groups to compare the means and the results (from the Pearson s Chi-square tests) of the sociodemographic attributes such as sex, marital status and education across the two groups. Secondly, they compared the means of the variables across the occupational groups and these variables included age, professional associations, social associations, journals read, network density, degree centralisation and information centrality. A statistical-significance test of the hypothesis that the means were not equal was also conducted (tstatistic). Thirdly, they cross-classified alter s relative rank and ego-alter tie strength against the occupational group to show (using Pearson s Chi-Square test) the statistical significances between the two groups. The measures of density, centrality and centralisation were then used to analyse the relational data. Some of the social network measures used are discussed in greater detail in the below section entitled Social Network Data Analysis. The study of the effects of social networks and ICT use on performance of rural GPs of NSW, Australia builds on the tradition established by Coleman, Katz et al [11] and the methodology utilised by West et al [44]. Burt s [10] proposal to General Social Survey (GSS) motivates the methodology of this study. Therefore, the study bases its survey instrument on questions provided by GSS, with a few modifications to suit the focus on social networks of rural GPs. To understand social network effects on performance, both relational and attribute data need to be collected and linked to facilitate analysis. Attribute data will include performance, ICT use, and personal attributes such as age, education, journals read, and memberships of professional and social associations of the GPs. Relational data includes elicitation of five alters with whom the GP (ego) discuss important professional matters within a certain time-frame (eg. past six months), detailed information about each alters, nature of relationship between the ego and alter, as well as the nature of relationship between alters. The dependent variable, performance, is classified into task based and contextual based performance [8]. Task-based performance includes those activities that are directly or indirectly related to the technical core of GPs activities. Contextual based performance is defined as behaviours that support broad organisational, social, and psychological environment of the organisation in contrast to behaviours that support the organisation s technical core. The use of ICT by GPs for their task or contextual-based activities, which impacts on any of the constructs of performance is a function of ICT use (an independent variable in this study). For example, decision support systems may be in place to assist the diagnosis of patient problems, and softwares may be deployed for suggesting or reminding GPs about patients medical prescription [20; 41]. Additionally, ICT is also considered to support and facilitate social networks by enabling actors to communicate in real time at low cost regardless of geographical location [43; 26; 29; 32]. Currently, there is lack of research evidence of rural GPs using ICT to seek advice or information from their social network. In Australian general practices, most ICTs are used for clinical and administrative activities, such as book-keeping, patient records, and prescriptions, rather than for communication [46]. Although most general practices are highly computerised, those in capital cities were less likely to be computerised than in rural practices [45]. The primary reason for this was because the rural GPs were geographically distant from the other health care centres, general practices, and hospitals, and therefore had a greater need for computerisation and electronic connectivity [23]. The implication is potentially significant in that GPs in rural NSW are ICT-enabled in various ways, but its use has not been measured to date. 5. GENERAL SAMPLING STRATEGY Similar to a social survey, sampling remains an important issue in network survey design especially when dealing with a large population. The population in this study is the GPs of rural NSW, Australia. The general practice serves as the sampling unit and the general practitioners (their ties) are the observation unit. There are 17 Divisions of General Practice within rural NSW, Australia totaling 1,518 GPs [1]. These divisions serve as a readily available mechanism for both the stratification and clustering of the sample of rural GPs. A simple estimation of the sample size at the 95% confidence level and at the 5% confidence interval shows that 307 GP responses are needed. The stratification sampling strategy presents two potential issues. Firstly, if the divisions are deemed to be the stratum, then we need to consider what constitutes an appropriate sub-sample within each division given the confidence interval and confidence level. Secondly, a simple random sampling strategy of all the 17 divisions is expensive in terms of time and money because the researcher would have to travel to all the 17 divisions across NSW in order to collect data. The main advantage of conducting a stratified sampling approach is that the data would be more representative of the rural NSW GPs because the data would have been collected from a random sample of all the 17 divisions. Further, recent research has shown that telephone interviews provide a more effective and valid measure to collect network data because of the caller s anonymity especially in sensitive settings [31]. Therefore, the issue of traveling becomes void because the collection of network data from GPs in rural NSW will be mostly conducted by telephone. The cluster sampling strategy remains a more viable option in terms of time and financial cost. Considering each division of general practice to be a cluster, one may choose a random sample of 7-10 divisions amongst the 17 and administer the survey to all the GPs of that division. The issue with this approach however, lies in the problem of non respondents. Although the ego-centric approach addresses this problem partially in that there is the chance that other GPs surveyed in the study might nominate the missing GP as one of his alters, the attribute data of the missing GP pertaining to ICT and performance may not be ascertained. This affects the reliability of the data collected as well as the extent to which it represents the population of rural NSW GPs. There is also a risk in that the divisions chosen as part of the 4

5 random sample may be very low in the population of GPs as compared to the other divisions which have a relatively higher population. This biases the sample and overall generalization of the data. The proposed solution to address the above concerns is to cluster the divisions of rural general practices by their demographic proximity into zones. Three zone emerge namely, the North East Zone (covering 8 divisions), the South East Zone (covering 6 division numbers), and the North West Zone (covering 3 divisions) in NSW. This strategy allows us to conduct a random sample of the rural divisions within each zone (treated as a stratum), thereby allowing us the advantage of conducting a random sample of GPs across a random sample of divisions across all the zones. A representative sample of the population is achieved in this way. Also, this strategy overcomes the constraint in the cluster sample strategy where each and every GP who falls within the division (here, treated as cluster) needs to be surveyed. 6. NETWORK SAMPLING ISSUES There are some general assumptions [28] that are taken into account in this study to collect social network data from rural GPs in NSW. First, the proportion of alters in a GP s social network are also members of the population. Second, it is assumed that everyone has an equal chance of knowing someone in a given population. Third, everyone has perfect knowledge about the members of their social network. Bernard et al [3] suggest that the proportion of alters elicited from the respondents allow for an approximation of the population. This is expressed as: m = c where m is the average number of alters that respondents know in the subpopulation, c is the average size of the respondent s network, e is the total size of the subpopulation and t is the total population. The problem with this equation to model sample size is that the assumptions mentioned above are usually not met for all subpopulations [30]. To account for this problem, Granovetter [22] provides a theoretical framework for which acceptable estimates for very large populations may be generated. The assumption that Granovetter makes is that the average acquaintance volume (V), (ie. the average number a respondent knows) can be computed from the calculation of a density of a network. Assuming that the ties are symmetric, and N t is the number of ties observed, and N is the number of actors in the network, density can be calculated as: 2 N t D = N ( N 1) Now, given that the N t ties observed represent two cases of an alter knowing other alters, then the total number of contacts in the network must 2N t and the average number per person is 2N t /N. The average acquaintance volume therefore is embedded in the calculation of density because V=(N-1)D by simple algebraic manipulation. Using this average acquaintance volume Granovetter then builds on Frank Ove s [16] work to mathematically compute the density estimate which includes the variance of the true vector of outdegrees (individual acquaintance volumes) and the variance e t of the true (population) sociomatrix. Through the calculation of the density estimate, Granovetter statistically proved that using a single large sample, it is feasible to estimate large populations while promising an unambiguous structural estimate with known variance. In fact, he proposes that it is theoretically and methodologically more sound to sample a few large samples rather than many small ones. The following graph demonstrates between the relationship of a given average acquaintance volume, the population size and the estimated sample size required: Figure 1: Sample size required for 95% confidence limits on network acquaintance volume, 20% error, for population sizes from 100-1,000,000, and true volume (V) of 100, 500, and 1,000. [22] Table 1: Table showing the sample sizes needed to meet 20% Error, 95% Confidence Limits, For w=1 (ie. No. of Sample(s)) [22] Using Granovetter s model confirms that the estimated random sample of approximately 300 GPs is sufficient. However, the formula for the density estimate requires knowledge of the population density and outdegree which is difficult to ascertain. This limits the practicality of the model. Moreover, the model assumes that the relational data collection strategy follows a whole network approach. This is because the rural GPs will be randomly sampled, a list of their names will be made up, and each GP would be asked whether he or she knows each individual name on the list. As mentioned earlier, in a whole network study, respondents are asked to identify a list of recognizable names from a given roster. Granovetter ([22] pg ) suggests that respondents can easily cope with 500 names from large populations. Other studies, however, show that about 150 is most feasible [34] [15]. 5

6 7. SOCIAL NETWORK DATA ANALYSIS Data gathered from the GPs will be stored in a mysql database. This will allow for flexible retrieval of both attribute and relational data to perform both descriptive statistics and social network analysis. The following section describes standard measures of social network data [39] that will be applied in the context of GPs : Network Density: Network density basically represents the actual number of ties in a network as a ratio of the total maximum ties that are possible with all the nodes of the network. A fully dense network has a network density value of 1, which indicates that all nodes are connected to each other. A network with a density value near 0 indicates that it is a sparsely-knit network. For an undirected graph with N nodes and N t ties the density D is defined as: 2 N t D = N ( N 1) Local Centrality and Global Centrality: Local centrality measures the number of direct ties that a particular node has, whereas global centrality measures indirect ties as well (i.e. ties that are not connected directly to that node). This said, a node that lies at a short distance between many other nodes is considered as close to many other nodes in the network (also termed as closeness ). Freeman [17] has proposed the measure of relative centrality to measure the centrality of a node with respect to the overall centrality of the other nodes in the network. His significant contribution in this field has enabled social network analysts to measure the node-centrality on a weighted basis which can be easily compared within the entire network. In mathematical terms degree centrality, d(i), of node i is defined as: d ( i) = m ij j where m ij =1 if there is a link between nodes i and j, and m ij = 0 if there is no such link. Centralisation: Centralisation and density are not only important measures in SNA, but they are also complementary to each other. Density explains the general level of connectedness in a network. Centralisation explains the extent to which the connectedness is focused around a particular node. To measure centralisation in a network, we need to observe the differences in the centrality values of the most central nodes and all the other nodes. Then, to arrive at the centralisation value, we calculate the ratio of the sum of actual differences and the sum of the maximum possible differences. Centralisation is thus defined as: g [max( D ) D ] i= 1 i i r = ( g 1)( g 2) where D i is the number of people in the network that are directly linked to person i. The number of actors are represented by g in this equation. Betweenness: Betweenness measures the extent to which a particular node lies in between the other nodes of the network. Betweenness centrality may be defined loosely as the number of times a node needs a given node to reach another node. Stated otherwise, it is the number of shortest paths that pass through a given node. As a mathematical expression the betweeness centrality of node i, denoted as b(i) is obtained as: g b( i) = g j, k where g jk is the number of shortest paths from node j to node k (j,k i), and g jk is the number of shortest paths from node j to node k passing through node i. 8. LIMITATIONS OF SOCIAL NETWORK ANALYSIS A real problem with network analysis in the past has been the inability to test hypotheses statistically, because the data are by their very nature auto-correlated, violating assumptions of independence (random sampling) built-in to most classical statistical tests [36]. With the advent of permutation tests, random graph models, and various multilevel models, however, this is much less of problem now [40; 35]. Furthermore, respondents are often asked to recall behaviour that took place over a broad period of time in order to capture as much information as possible. Brewer [9] has found that forgetting to name people in recallbased elicitation of sociocentric networks is a potentially significant problem when collecting such data. Also, if the time period is too long, or the amount of information too detailed, reliability and accuracy are jeopardised. Some social network analysts express concern that data based on recall, although widely used, may be less reliable than data gathered by observation. In light of this, the quality of social network data based on recall have been systematically studied [5; 6; 4] where it has been claimed that people are generally very inaccurate in reporting on their past interaction with other people. Later studies also confirmed the same findings [19; 12] but added that people also remembered long-term or typical patterns of interaction with other people rather well. Data gathered by self-reporting could explore the influence of other personal variables that cannot be explored when data is gathered by observation. For instance, the free recall method elicits a richer data on the social networks of people whereas the fixed choice method influences people to elicit accurate information on the most important relationships (ie. strong ties) [24]. Thus, recall may be better for understanding participants way of thinking, while observation may be better for understanding participants actual behaviour, but both are not mutually exclusive [21]. 9. SUMMARY This paper provides an overview of the concept of social network analysis in terms of understanding the performance of rural general practitioners. SNA has gained significant recognition as a methodology in recent years. Combining relational data with attribute data in the social survey, the insights gained from the analyses are richer than the traditional social survey. The two main approaches used to study social networks are sociocentric and egocentric network approaches. While the sociocentric approach is used to study a social network where the boundaries are specified clearly, the egocentric approach is gaining popularity because of its focus on individuals, groups and communities. This approach was utilized in a famous study on the adoption of innovation of a particular drug use by doctors. It was also used to understand the differences in the personal and social characteristics of doctors and nurses in a study in UK. The methods used in these two famous studies drive the methodological design for this study. In the end, issues related to network sampling, data collection and analyses are discussed. jik jk 6

7 10. REFERENCES [1] AGDP, "Australian Divisions of General Practice - Divisions Directory" &state_code=nsw (2nd February, 2005). [2] Anklam, P., "KM and the Social Network" the%20social%20network.pdf (11th February, 2005). [3] Bernard, H. R., Johnsen, E. C. and Robinson, S. (1989). Estimating the Size of an Average Personal Network and of an Event Subpopulation. The Small World. M. Kochen. Norwood, NJ, Ablex: [4] Bernard, H. R., Killworth, P. D., D, K. and L, S. (1985). "On the Validity of Retrospective Data". Annual Review of Anthropology 13: [5] Bernard, H. R., Killworth, P. D. and Sailer, L. (1982). "Informant Accuracy in Social-Network Data V. An Experimental Attempt to Predict Actual Communication from Recall Data". Social Science Research 11: [6] Bernard, H. R., Shelley, G. A. and Killworth, P. D. (1982). "How Much of a Network does the GSS and RSW Dredge Up?" Social Networks 9(49-61). [7] Borgatti, S. P. and Foster, P. C. (2003). "The Network Paradigm in Organizational Research: A Review and Typology". Journal of Management 29(6): [8] Borman, W. C. and Motowidlo, S. J. (1993). Expanding the Criterion Domain to Include Elements of Contextual Performance. Personnel Selection in Organizations. N. Schmitt and W. C. Borman. San Francisco, CA, Jossey Bass: [9] Brewer, D. D. (2000). "Forgetting in the Recall-based Elicitation of Personal and Social Networks". Social Networks 22: [10] Burt, R. (1984). "Network Items and the General Social Survey". Social Networks 6(4): [11] Coleman, J. S., Katz, E. and Menzel, H. (1957). "The Diffusion of an Innovation among Physicians". Sociometry 20(4): [12] Corman, S. R. and Bradford, L. (1993). "Situational Effects on the Accuracy of Self-Reported Organizational Communication". Communication Research 20: [13] Cross, R., Borgatti, S. P. and Parker, A. (2002). "Making Invisible Work Visible: Using Social Network Analysis to Support Strategic Collaboration". California Management Review 44(2): [14] Emirbayer, M. (1997). "Manifesto for a Relational Sociology". The American Journal of Sociology 103(2): [15] Erickson, B. H., Nosanchuk, T. A. and Lee, E. (1981). "Network Sampling in Practice: Some Second Steps". Social Networks 3: [16] Frank, O. (1971). Statistical Inference in Graphs. Stockholm, FOA Repro Försvarets Forskningsanstalk. [17] Freeman, L. C. (1978). "Centrality in Social Networks: Conceptual Clarification". Social Networks 1(3): [18] Freeman, L. C. (2004). The Development of Social Network Analysis: A study in the Sociology of Science. Vancouver, Empirical Press. [19] Freeman, L. C., K, R. A. and C, F. S. (1987). "Cognitive Structure and Informant Accuracy". American Anthropologist 89: [20] Garg, A. X., Adhikari, N. K. J., McDonald, H., Rosas- Arellano, M. P., et al. (2005). "Effects of Computerized Clinical Decision Support systems on Practitioner Performance and Patient Outcomes: A Systematic Review". Journal of the American Medical Association 293(10): [21] Garton, L., Haythornthwaite, C. A. and Wellman, B. (1997). "Studying Online Social Networks". Journal of Computer Mediated Communication 3(1). [22] Granovetter, M. (1976). "Network Sampling: Some First Steps". The American Journal of Sociology 18(6): [23] Group, G. P. C., General Practice Computing Group, (2002). Preliminary Report to GPRAC Rural and Remote Standing Committee on Uptake and Utilisation of IM/IT in Rural General Practice, Canberra. [24] Hammer, M. (1984). "Explorations into the Meaning of Social Network Interview Data". Social Networks 6: [25] Hanneman, R. A., "Introduction to Social Network Methods" (8th August, 2004). [26] Haythornthwaite, C. (2002). "Strong, Weak, and Latent Ties and the Impact of New Media". The Information Society 18: [27] Hirdes, J. P. and Scott, K. A. (1998). "Social Relations in a Chronic Care Hospital: A Whole Network Study of Patients, Family and Employees". Social Networks 20: [28] Jackson, D., Kirkland, J., Jackson, B. and Bimler, D. (2005). "Social Network Analysis and Estimating Size of Hard-to- Count Subpopulations". Connections 26(2): [29] Katz, J. E. and Rice, R. E. (2002). Social Consequences of Internet Use: Access, Involvement, and Interaction. London, MIT Press. 1st Edition. [30] Killworth, P. D., Johnsen, E. C., McCarty, C., Shelley, G. A., et al. (1998). "A Social Network Approach to Estimating Seroprevalence in the United States". Social Networks 20: [31] Kogovšek, T. and Ferligoj, A. (2005). "Effects on Reliability and Validity of Egocentered Network Measurements". Social Networks 27(3):

8 [32] Licoppe, C. and Smoreda, S. (2005). "Are Social Networks Technologically Embedded? How Networks are Changing Today with Changes in Communication Technology". Social Networks Volume: In Print. [33] Marsden, P. and Campbell, K. E. (1984). "Measuring Tie Strength". Social Forces 63: [34] Morgan, D. L. and Rytina, S. (1977). "Comment on 'Network Sampling: Some First Steps' by Mark Granovetter". American Journal of Sociology 83( ). [35] Morris, M. (2003). Local Rules and Global Properties: Modeling the Emergence of Network Structure. Dynamic Social Network Modeling and Analysis. R. Breiger, K. Carley and P. Pattison. Washington, DC, National Academy Press. [36] Morris, M. (2004). Overview of Network Survey Designs. Network Epidemiology. M. Morris. Washington, OUP. [37] Otte, E. and Rousseau, R. (2002). "Social Network Analysis: A Powerful Strategy, also for the Information Sciences". Journal of Information Science 28(6): [38] Scott, J. (2000). Social Network Analysis: A Handbook. London, SAGE Publications. [39] Wasserman, S. and Faust, K. (1994). Social Network Analysis: Methods and Applications. New York, Cambridge University Press. [40] Wasserman, S. and Pattison, P. (1996). "Logit Models and Logistic Regressions for Social Networks: I. An Introduction to Markov Graphs and p*". Psychometrika 60: [41] Wears, R. L. and Berg, M. (2005). "Computer Technology and Clinical Work: Still Waiting for Godot". Journal of the American Medical Association 293(10): [42] Wellman, B. (1926). "The School Child's Choice of Companions". Journal of Educational Research 14: [43] Wellman, B. (1992). "Which Types of Ties and Networks Give What Kinds of Social Support?" Advances in Group Processes 9: [44] West, E., Barron, D. N., Dowsett, J. and Newton, J. (1999). "Hierarchies and cliques in the social networks of health care professionals: implications for the design of dissemination strategies". Social Science and Medicine 48: [45] Western, M. C., Dwan, K. M., Western, J. S., Makkai, T., et al. (2003). "Computerisation in Australian General Practice". Australian Family Physician 32(3). [46] White, C., Sheedy, V. and Lawrence, N. (2002). "Patterns Of Computer Usage Among Medical Practitioners In Rural And Remote Queensland". Australian Journal of Rural Health 10: