A Cautionary Note on Data Inputs and Visual Outputs in Social Network Analysis

Innovations in network visualization software over the last decade or so have been important to the popularization of social network analysis (SNA) among academics, consultants and managers. Indeed, there is a growing literature that seeks to demonstrate how ‘invisible social networks’ might be revealed and leveraged for ‘visible results’ through management interventions. However, the seductive power of the network graphic has distracted attention away from a variety of emerging and long recognized concerns in SNA. For example, weaknesses exist in data collection techniques that often rely on nominal boundary‐setting and respondent recall. Non‐response can also be highly problematic. Increasingly, email data are being employed, yet this represents a poor proxy for relationships and raises issues of privacy. In displaying relational data, visualizations typically reify and ossify the network. Yet, individual perceptions of a network can vary greatly from unified visualizations, and their structure is typically fleeting. The aim of this paper is to draw together the diffuse literature concerning data input and visual output issues in SNA, in order to raise awareness among management researchers and practitioners. In doing so, the nature and impact of such weaknesses are discussed, as are ways in which these might be resolved or mitigated.


Introduction
The network literature has grown exponentially in recent years across a wide range of fields, including business and management (Borgatti and Foster, 2003, p. 992). A key approach adopted in this literature is that of social network analysis (SNA) (e.g. Ahuja and Carley, 1999;Allen, James and Gamlen, 2007;Cantner and Graf, 2006;Casper, 2007;Cattani and Ferriani, 2008;Cross, Borgatti and Parker, 2002;Kijkuit and van den Ende, 2010). It is argued that the emergence over the last 10-15 years of powerful and freely available network visualization tools (e.g. Krackplot, 1 UCINET, 2 Payek, 3 Metasight 4 ) has encouraged the use of SNA techniques by management academics, and fuelled their popularization among business consultants and managers. 5 Indeed, there is a growing literature that seeks to demonstrate how 'invisible social networks' might be revealed Early drafts of this paper were presented at a departmental seminar at the University of Leicester School of Management and at the 28th SCOS conference in Lille, France. I am very grateful for the constructive comments and feedback from colleagues at both of these presentations as well as from two anonymous reviewers, which have helped shape the ideas and focus of this paper.  Freeman (2000) for an overview of the history and diversity of social network visualization tools, and see wikipedia.org/wiki/Social_network_analysis_software (accessed 22 November 2011) for a good overview of a large range of software applications for the visualization of social network data and links to websites for individual applications. and leveraged for 'visible results' within organizations (Cross and Parker, 2004;Cross and Thomas, 2009;Cross, Borgatti and Parker, 2002;Krackhardt and Hanson, 1993). Whilst it is recognized that not all network research employs visualization tools to depict the social structure under investigation, there are nevertheless a rapidly growing number of examples that can be found within the academic literature, including the majority of the studies referenced in this paper, as well as in practitioner texts and on consultancy websites.
However, despite the growing use of SNA by business and management academics and practitioners, it is contended that too little attention in the literature has been focused on the nature of the data being collected, the manner in which it is being displayed, or the associated ethical issues in such studies. For example, it is common for SNA studies in the management field to be silent or underplay important issues relating to boundary-setting, informant response rates and decisions concerning network visualization (e.g. Allen, James and Gamlen, 2007;Chiffoleau, 2005;Stephenson and Lewin, 1996). Ethical issues in relation to SNA research are raised rarely. The objective of this paper then is to heighten awareness of these concerns within the business and management community. Issues concerning individual techniques for processing and analysing social network data tend to be highly technical and, as such, are considered better dealt with in the specialist social network literature. 6 In this paper we start by providing an overview of the scope of SNA usage across the field of business and management. We then turn to an evaluation of the accuracy and completeness of the data in such network studies, and highlight possible ways in which weaknesses apparent in survey methods, for example, might be mitigated. We consider the nature of the network visualization itself, reflecting on the multiple ways in which a network may be viewed and depicted and how such depictions may be interpreted. Finally, we surface the ethical and privacy issues associated with network research. These are increasingly pertinent because of the rise in use of SNA by consultants and managers in relation to decisionmaking within organizations Parker, Cross and Walsh, 2001). Indeed, Borgatti and Molina (2003, p. 338) rightly warn us that 'consideration of ethical issues [is] increasingly critical as organizations start basing personnel and reorganization decisions on network analyses'.

The breadth of SNA usage in business and management
Over the last couple of decades there has been a rapid growth in the use of SNA techniques to research a wide range of business and management issues and contexts. More recently, such techniques have been applied to the study of specialist academic communities within business and management itself. However, perhaps most interesting is its diffusion into business consultancy and business practice.
There are also a growing number of fascinating SNA studies that have turned the gaze inward, onto the academic communities within business and management, such as those mapping the 6 Such as Social Networks, Sociometry, Connections, the Journal of Social Structure and the Journal of Quantitative Anthropology. Data Inputs and Visual Outputs in Social Network Analysis 103 'invisible college' among B2B marketing researchers (Morlacchi, Wilkinson and Young, 2005, p. 14), economists concerned with technology and innovation (Verspagen and Werker, 2003, p. 408;, p. 1425, the information systems community (Vidgen, Henneberg and Naudé, 2007), hospitality management researchers (Hu and Racherla, 2008, p. 306) and around specific journals, such as R&D Management (McMillan, 2008, pp. 74-76). Broader based studies have also sought to map the invisible college among the most prominent researchers in management and organization studies (Acedo et al., 2006, pp. 976-977), and the interconnectedness of editorial board membership across the Financial Times 40 management and business journals (Burgess and Shaw, 2010, pp. 636-640). Hu and Racherla (2008), for example, as with a number of the above studies, employ co-authorship data from prominent journals in the field. Whilst they recognize limitations to their study, such as its inability to capture informal interactions, they suggest worryingly that the resulting network maps could 'serve as alternative metrics to evaluate (or at least imply) research impacts and contributions of individual researchers by research collaborations, which in many cases is difficult to detect by the conventional methods' (Hu and Racherla, 2008, p. 311). Over the last decade, SNA techniques have also been applied increasingly in consultancy work, in order to reveal informal structures and knowledge flows and identify influential individuals, such as gatekeepers and opinion leaders. Cross et al. (2001) argue that SNA achieves this by enabling the production of an 'X-ray' of the informal network. Parker, Cross and Walsh (2001), for example, have applied such techniques within a consortium of Fortune 500 companies and government agencies, often as a precursor to identifying 'intervention opportunities'. In one case, involving a consulting practice, Parker, Cross and Walsh (2001, p. 27) argue that 'the result of interventions was significant . . . the group began to sell more . . . [and] a network analysis conducted nine months later revealed a well-integrated group that was leveraging and seeking its knowledge much more effectively'.
This consultancy work reflects a growing recognition within the areas of human resource management and organizational development of the potential of SNA (Bunker, Alban and Lewicki, 2004;Hatala, 2006;Lengnick-Hall and Lengnick-Hall, 2003;Stephenson and Lewin, 1996). Indeed, Hatala (2006, p. 65) argues that 'SNA can provide HRD [human resource development] practitioners with valuable relational information that can assist in the assessment of performance and implementation interventions'. However, despite their extensive research and consulting work, Parker, Cross and Walsh (2001, p. 28) recognize that 'network analysis is not a cure-all' and that 'if applied without proper forethought, the results can be inconclusive at best and damaging at worst'. This point is important to reflect upon since, as Borgatti and Molina (2003, pp. 337-338) stress, 'The stakes are higher in the practice setting than in the academic setting, because the purpose of the network research there is explicitly to make decisions that directly or indirectly will affect the lives of employees'.

Evaluating the accuracy and completeness of SNA data inputs
The nature of network data Social networks comprise three core components: actors, links, and flows. They are constructed by identifying and then connecting individual dyads. Typically, such network data are obtained through a questionnaire survey completed by the members of the network, 7 although data can also be collected through interviews, documents, observation and from various electronic sources. A link is considered to exist where both actors in a dyad report a relationship with the other; this is termed 'reciprocal nomination' (Stork and Richards, 1992). However, 'non-reciprocal nominations' may be 'symmetrized' (Scott, 1991), i.e. a relationship may be considered to be present even when it is reported by only one of the two individuals in the dyad.
Since network data cannot typically be collected instantaneously and may relate to an event taking place over a period of time (e.g. the relationships mobilized during the development of a 7 For a sample of social network questionnaires see the webpage for Professor Krackhardt of the Heinz School of Public Policy and Management, Carnegie Mellon University (www.andrew.cmu.edu/user/krack/ questionnaires.shtml).

104
S. Conway new product), it is subject to 'temporal grouping'. That is, network data are aggregated from across the period of data collection, effectively conflating time and disregarding the ordering of relational events. Collecting network data from blogs, newsgroups, email and chat rooms is becoming more common and may serve to resolve some of these issues, although such internet sources present their own 'accuracy' and ethical issues. It has been argued that social network studies often under-emphasize the flows through the network, whilst over-emphasizing the quantity rather than the 'quality' or 'utility' of network relationships and interactions (Conway, Jones and Steward, 2001).

Problems associated with boundary-setting and choices concerning 'rules of inclusion'
Where SNA is being undertaken among an identifiable group of individuals, such as a project team or department within an organization, then the membership is likely to be reasonably clear to the researcher. Nevertheless, it is often the boundary-spanning relationships that are of particular interest and importance to researchers and managers alike (Tushman, 1977), and these linkages can be remain 'hidden' if the boundary around data collection is set too tightly to the membership of the group. However, in many cases the membership of the group of individuals under investigation is poorly defined, such as with informal networks and communities (Ghani, Donnelly and Garnett, 1998). In such instances, the researcher may sensibly begin by approaching those known members and proceed by identifying further members from the responses from these known members (Scott, 2000, p. 61). In employing such a 'snowball' sampling approach, the network researcher must at some point decide where and when to stop collecting data, otherwise they will be drawn into 'the general everramifying, ever-reticulating set of linkages that stretches within and beyond the confines of any community or organisation' (Mitchell, 1969, p. 12). Yet, in doing so, the researcher sets a nominal boundary for the network and effectively decides who is, and therefore who is not, part of the network (Laumann, Marsden and Prensky, 1983).
In very large bounded groups, such as a company division of several hundred employees, the collection of network data can very quickly become unmanageable. In such instances, it is not advisable in SNA to simply select a representative sample, since this does not provide a 'useful sample of relations' (Scott, 2000, p. 59). One strategy to cope with large networks is for the researcher to establish their own 'rules of inclusion', which may be based on characteristics such as the role, seniority or gender, for example, of the members of the larger group. Such rules of inclusion should be clearly linked to the research questions of the project (Laumann, Marsden and Prensky, 1983).
A specific example of these boundary-setting and sampling decisions can be seen in a recent investigation of the informal problem-solving network within ICI's R&D function (Allen, James and Gamlen, 2007). In this study, to make data collection manageable, only senior personnel were selected, representing 152 of approximately 400 R&D staff. Furthermore, the researchers did not look at interactions across the organizational boundary, although they recognized that 'external networks and links with scientific communities are very important for research scientists' (Allen, James and Gamlen, 2007, p. 184). As a result, the informal problem-solving network that is identified by the researchers is partial, and underrepresents the complexity and diversity of the internal and external linkages. Thus, boundarysetting and sampling decisions can have a profound impact on the structure of the network that is revealed, and as a result Fombrun (1982, p. 288) warns that the conclusions drawn from a network study 'must be carefully scrutinized for the possibility of alternative explanations grounded in the effects of the untapped networks'.

Problems associated with missing or inaccurate data
Having established the boundary and rules of inclusion for a network study, it is important that close to a 'complete' data set is obtained by the researcher. Parker, Cross and Walsh (2001, p. 28) argue that 'while you don't have to get 100% response, we typically shoot for at least 80% response from the group we're analysing'. In contrast, others contend that 'the analysis and mapping of the structure of the network is especially sensitive to missing data' (Huisman, 2009, p. 2) and that missing data can be 'very misleading  (Borgatti and Molina, 2003, p. 339). The latter point is emphasized graphically in Figures 1 and 2, which illustrate the distortion of the network structure as a result of missing data (inspired by Borgatti and Molina, 2003, p. 340). In Figure 1, data have been collected and mapped for all actors and relationships in a network, whilst in Figure 2, depicting the same network, data are missing for two actors, i.e. actors 10 and 11, who hold impor-tant positions in the network. As a result of these missing data in the latter visualization, the bridge between the two networks remains invisible to the researcher. Such bridges are considered important for promoting novelty and creating entrepreneurial opportunities (Burt, 1992, p. 26).
Missing and inaccurate network data can arise from a number of sources. Principal among these are the non-response of network members, questionnaire design, and informant bias (Kossinets, 2006).

Missing data arising from the non-response of network members
The problems in collecting network data are often compounded by the non-response of a proportion of network members. Questions typically employed in collecting network data are 'sensitive' (Tourangeau, Rips and Rasinki, 2000, p. 255), and the mapping of network data can expose the network status of individuals. This may further deter individuals from being involved in such research, especially where it is being employed to make managerial decisions (Hatala, 2006). To a certain extent, non-responses can be ameliorated through a process of 'symmetrization' (Scott, 1991). That is, where a network member does not respond to a survey, it might be possible to determine their connections where network members that do respond indicate links with these non-respondents. Clearly, the efficacy of such an approach diminishes as the percentage of non-response increases, although simulations indicate that reasonable results can be achieved with up to a 20% non-response rate (Huisman, 2009). Even so, response rates below 100% have the potential to miss crucial network linkages.

Missing data arising from questionnaire design
Questionnaires are a common tool for collecting SNA data, and thus questionnaire design also plays its part in the 'completeness' and 'reliability' of a network data set. SNA questionnaires typically incorporate only a very limited number of questions since these often need to be answered in relation to a sizable group of individuals. The questionnaire may include the full list of names of the group under investigation, against which respondents may be asked to confirm all those individuals with whom they communicate. However, this technique is not possible where the membership of the group is not clear to the researcher. In such circumstances the questionnaire may be employed to reveal the network membership by asking the known members to indicate the names of those with whom they communicate. Such an approach would then employ a snowball sampling strategy. In both instances, recall by respondents of weak connections or infrequent interactions can be an issue, and may be compounded where the group is particularly large or where the full names of contacts are not known by respondents. This is important since, whilst strong ties promote information flow, weak ties provide information novelty (Burt, 1992, p. 26). It is also important when designing questionnaires to be wary of the terms employed. For example, many network studies ask respondents to identify their friends, yet even the term 'friends' is very ambiguous and can mean different things to different respondents (Fischer, 1982).

Inaccurate data arising from informant bias
Following a number of experiments to test informant accuracy in reporting past communications, Bernard et al. (1984, p. 499) concluded that 'what people say about their communications bears no useful resemblance to their behavior', since respondents recalled less than 50% of their interactions correctly. They found that respondents typically make two types of recall error -they forget some of those with whom they have interacted and they incorrectly recall interactions with others with whom they have not. In addition to general 'memory decay', there are a number of factors that impact the accurate recall of interactions and relationships, such as their perceived salience by the respondent, the specificity of the behaviour being investigated and the size of the network (Bell, Belli-McQueen and Haider, 2007). Furthermore, and not surprisingly, respondents are much better at recalling their own relationships (i.e. 'direct' or 'first-order' connections) than the relationships of those with whom they are connected (i.e. 'indirect' or 'second-order' connections). Such research would seem to undermine dramatically the utility of 'recalled' network data. However, subsequent research by Freeman, Romney and Freeman (1987, pp. 321-322) found that informants typically drew from 'somewhere between experience and recall' in their responses. That is, 'what is recalled . . . is what is typicalwhether it happened or not'. Interestingly, as a result, this research reveals that individuals are, in fact, very good at recalling enduring patterns of relations with others, although this will lead to an under-reporting of weak ties. The accuracy of network data may also be distorted by 'selfpresentation' (Goffman, 1973), i.e. respondents perhaps wanting to be viewed as more connected or interactive than they actually are. Research concerning the self-presentation of individuals on Data Inputs and Visual Outputs in Social Network Analysis 107 social networking sites, for example, has found that users often present 'hoped-for possible selves' online that differ from their 'real selves' offline (Zhao, Grasmuck and Martin, 2008).

Alternative data collection methods and alternative data sources
Although network studies are often associated with the collection of data via questionnaires, a variety of methods and data sources may be employed to reveal network data. These include, for example, interviews (e.g. Cross et al., 2001;Freeman, Freeman and Michaelson, 1989), observation (e.g. Conti and Doreian, 2010;Freeman, Freeman and Michaelson, 1989), biographies (Crossley, 2008), personal letters (Edwards and Crossley, 2009), co-citation data from journal articles (e.g. Acedo et al., 2006;Hu and Racherla, 2008) and social networking sites, such as Facebook (e.g. Lewis et al., 2008). Each approach has its inherent strengths and weaknesses. Lewis et al. (2008, p. 341), for example, highlight the ease with which network data can be obtained from Facebook, whilst also recognizing that respondents 'differ tremendously in the extent to which they "act out their social lives" on Facebook'. There is evidence that network researchers are increasingly employing multiple methods in order to yield complementary data (e.g. Conti and Doreian, 2010;Crossley, 2008;Edwards and Crossley, 2009;Human and Provan, 1997;Park and Kluver, 2009). In this regard, quantitative approaches may be viewed as being relatively effective at revealing the structure of networks, whilst the in-depth data available through qualitative approaches may be seen as more effective in providing insight into the process, content and context of relationships and interactions. In some cases mixed methods have been employed explicitly to triangulate the data. Lievrouw et al. (1987), for example, in their study of the intellectual connections between biomedical scientists, employed both co-citation data and interviews. Nevertheless, in recent years there has been an increasing recognition that qualitative methods have been under-utilized and that there is a need for the adoption of mixed methods in network research to broaden and deepen our understanding (Coviello, 2005;Hoang and Antoncic, 2003;Jack, 2010). In particular, Coviello (2005) argues that mixed methods have a useful role to play in collecting data on network dynamics.

Network visualization and early visualization techniques
Academic studies have employed network visualization techniques for over 75 years to reveal the social structure in a huge variety of interesting contexts, from mapping the social structure among cohorts of school pupils (Moreno, 1934, pp. 154-161), the interlocking directorates between organizations (Levine, 1972, p. 15) and the spread of AIDS through social contacts (Klovdahl, 1985(Klovdahl, , p. 1204, to revealing terrorist networks (Krebs, 2002, pp. 46, 50), informal connections in Formula 1 (Henry and Pinch, 2000, p. 200) and the social network of the UK 'punk' movement (Crossley, 2008, p. 101). Moreno (1934) is generally credited with the first attempts to visualize social networks. His 'sociograms' were hand-drawn depictions. So too were the network visualizations of others in the subsequent decades (e.g. Levine, 1972;Roethlisberger and Dickson, 1939, pp. 502-507). Yet, despite the power of the graphic for displaying relational data, network visualizations remained relatively rare until recently. Klovdahl (1986, p. 313) attributes this under-utilization to 'the time and tedium involved in producing hand-drawn diagrams' and 'the impossibility of manipulating these once they are drawn'. During this period, a 'data matrix' was widely employed to record and display network data (Scott, 2000, p. 40). Table 1 represents the network in Figure 3 as a data matrix, where a '1' indicates that a link exists between two actors and a '0' indicates that no link exists. Interpreting such matrices remains a skill associated with experienced network researchers. Given these alternatives for displaying network data, it is easy to see why the emergence of network visualization software has been such an important innovation in the popularization of SNA, particularly among consultants and practitioners.
One of the key features of network visualization software is the ease with which it allows the researcher to manipulate the graphic, such as in the re-positioning or removal of actors. Many SNA software packages allow for the automated presentation of network data using what is termed 'multi-dimensional scaling'. Scott (2000, p. 149)

108
S. Conway notes that at its simplest multi-dimensional scaling is a technique for converting network metrics, such as 'centrality' and 'path distance', into physical distance on the screen or page. This can be a powerful way for revealing clusters, for example. However, as the values of network metrics change, so too do the physical positions of individual actors on the screen or page, which can be confusing when attempting to compare a network at different points in time.
Network visualizations are currently employed in a number of different ways. As an output from a network study, they can provide a powerful medium for displaying and revealing the key features of the network under investigation, such as 'clusters', 'structural holes' and 'bridges'. These, in turn, can inform consultants and practitioners of potential interventions to alter the morphology of the network toward particular goals, such as improving communication and knowledge flow between distinct organizational groups. Network maps are sometimes used during the data collection process itself, for example, as a way of interacting with respondents to confirm the 'completeness' of a network created from an earlier data collection phase. They can also be employed to co-create the network in 'real time', a process sometimes referred to as 'participatory mapping' (Lubbers et al., 2010), and they have been employed to help guide the researcher toward fruitful areas of focus during subsequent data collection phases (Biddex and Park, 2008;Park and Kluver, 2009). Network visualizations also have a role to play in aiding the process of theory building, since through the manipulation of a depiction new insights can emerge (Conway and Steward, 1998;Klovdahl, 1981Klovdahl, , 1986Moody, McFarland and Bender-deMoll, 2005).

'The map is not the territory': the multiple visual representations of a network structure
The role of the researcher in designing the network depiction It is clear from a review of a broad array of network depictions in the literature, such as those indicated earlier, that there is considerable variety in the network data that are displayed and the way that these data are represented. This is perhaps not surprising, given that it would appear that many network visualizations are arrived at through trial and error (Bertin, 1983, p. 271;Freeman, 2000) without recourse to 'a set of recognised conventions' (Bender-deMoll and McFarland, 2006;Conway and Steward, 1998). Indeed, there is no 'one right way' to depict a network (McGrath and Blythe, 2004;Scott, 2000, p. 65). There are two prominent approaches to the production of network graphics. The first might be labelled a 'graphical excellence' approach, as typified by the work of Tufte (1983Tufte ( , 1990 and Bertin (1983). From this orientation, 'excellence in . . . graphics consists of complex ideas communicated with clarity, precision, and efficiency' (Tufte, 1983, p. 13). This is achieved through the considered use of what Bertin (1983, p. 71) has termed the 'visual' or 'retinal' variables, such as size, colour and shape, in depicting the individual actors, links and flows. The second may be termed a 'visual argument' approach (Simon, 1969, p. 5). From such a standpoint, Levine (1972, p. 14) argues that 'the value (or deceptiveness) of a [graphical] representation lies in what it suggests . . . its ability to stimulate thought'. Whilst these two approaches are potentially complementary, for Tufte (1983, p. 51) 'graphical excellence' requires the researcher to 'tell the truth about the data' via the visual display; clearly this is at odds with a perspective that seeks to emphasize a particular version of the 'truth'.
There is often a trade-off between 'seeing the overall forest -the clusters of overall groups and their relative social proximity or ordering in relation to each other -and seeing the finer detail of the trees -identifying key players and roles within these groups' (McGrath, Krackhardt and Blythe, 2003, p. 46). Through the graphic, the researcher may seek to highlight particular features of the network, such as clusters of actors, bridges between clusters, or the diversity and size of the overall network (Bender-deMoll and McFarland, 2006;Conway and Steward, 1998;McGrath and Blythe, 2004). However, different spatial arrangements of the same network might either highlight or obscure such network features. This point is emphasized in Figures 3 and 4
Consideration must also be given to the selection of the characteristics of the actors, links and flows to be displayed. A network map is able to incorporate a variety of quantitative and qualitative information. However, the choice of visual variable to be employed in the display of such different types of data is crucial (Bertin, 1983, p. 71). Typically, 'size' is most effectively mobilized for the 'quantitative' features of actors and links, such as years of experience or the strength of a relationship. In contrast, 'colour' and 'shape' are best suited to displaying 'qualitative' features, such as an actor's gender or functional location, and the type of 'flow' through a link (e.g. knowledge, friendship, power).

The viewer's interpretation of a network depiction
There are a number of dangers in the choices made by researchers to encode certain features of the actors, links or flows, or in the manipulation of the network graphic in order to present a specific 'visual argument'. First, there is the possibility, whether intentional or accidental, that the viewer might be misled about certain characteristics of the network (Bender-deMoll and McFarland, 2006). Second, relatively little is known about how viewers interpret or decode the network visualizations they are presented with (Bender-deMoll and McFarland, 2006;McGrath and Blythe, 2004;McGrath, Krackhardt and Blythe, 2003). In part, this is because viewers 'bring a rich vocabulary of graphical idioms and conventions to the table when they interpret the visualization' (McGrath and Blythe, 2004, p. 1).

From researcher generated aggregated network maps to individualized 'cognitive maps'
It is the norm for network analysts to aggregate the network data of individual respondents to create a single network map. Yet there has long been evidence to indicate that individuals within a network may have very different 'cognitive maps' or 'cognitive structures' of the very same network (Krackhardt, 1987(Krackhardt, , 1990. That is, 'to some extent, social structure is in the eye of the beholder' (Kilduff and Krackhardt, 1994, p. 87). Colville and Pye (2010, p. 378) contend that 'this poses problems of aggregation . . . as you raise the level of the analysis from the individual to the collective in search of network insight'. Interestingly, a recent study by Kilduff et al. (2008) revealed that individuals perceive more clustering than is present in the 'actual' network and attrib-uted more popularity and brokerage to individuals they perceived as popular. As a result, Kilduff et al. (2008, p. 25) argue that: Perceiving the organization as a small world may reassure the individual concerning the approachability of even distant people . . . . On the other hand, a tendency to misperceive clustering . . . together with a tendency to attribute more importance to perceivedly popular people, may lead active networkers to be overly confident in picking key people in the network with whom to form attachments. Managers, for example, might assume that they are keeping in touch with all the important clusters, when, in fact, the clustering and connectivity they perceive are more figments of their imagination than accurate features of the social network.
That individuals perceive such different network structures may be seen as a process of 'sense-making' by individuals within an organization when they are faced with complex and dynamic webs of relationships (Colville and Pye, 2010;Purchase, Lowe and Ellis, 2010;Ramos and Ford, 2011;Weick, 1995, pp. 38-39). Thus, we might expect the cognitive map of an individual to shape their behaviour within a network rather than an aggregated version of the network map as represented by the researcher.

Emerging approaches to depicting networks -moving from 'snapshots' to 'movies'
Social networks are typically dynamic structures. However, attempts to 'capture' and 'make visible' networks have often led to the mapping of a single 'snapshot in time' of the network structure. In doing so, there is a danger that the network visualization presents an ossified version of the network. This is likely to reinforce the prevailing attention on 'static structures' rather than 'the dynamic processes that transform those matrices of transactions in some fashion' (Emirbayer, 1997, p. 305). Attempts to address this concern have led to a growing interest in longitudinal research in the study of social networks. Interesting examples include that of social network formation and inter-firm mobility within the San Diego biotechnology cluster (Casper, 2007), field evolution in the life sciences (Powell et al., 2005) and changes in managerial sense-making (Öberg, Henneberg and Mouzas, 2007).
Data Inputs and Visual Outputs in Social Network Analysis 111 In longitudinal research, data are typically collected at intervals and displayed as a series of snapshots (e.g. Casper, 2007;Degenne and Lebeaux, 2005;Powell et al., 2005). Yet for Moody, McFarland and Bender-deMoll (2005, p. 1207), such depictions 'do a poor job of representing change in networks', since whilst, on the one hand, longitudinal data might capture the enduring patterns within a network, on the other, the fluctuations in relationships and interactions between the sampling periods are lost. Furthermore, each snapshot suffers from the 'temporal grouping' noted earlier. Increasing the frequency of these discrete waves of data collection and the resulting number of snapshots can help mitigate these concerns, although this is likely to have a major impact on the effort required to collect the requisite data. Nevertheless, Bender-deMoll and McFarland (2006) argue that whilst we can 'talk usefully about network change [in such research] . . . it is difficult to argue that "dynamics" and "evolution" have been recorded'.
For researchers to effectively capture the dynamics of a network, they will need to 'tease apart' the relationship between the micro-level interactions and the overall network. Ideally, this would be done by capturing changes or activity as it occurs, to collect a continuous 'stream' of data. These data could then be displayed not as a series of discrete network pictures but as an animated 'network movie', with gradual changes in individual actors, links and flows that seamlessly and gradually reshape the network map (Bender-deMoll and McFarland, 2006;Moody, McFarland and Bender-deMoll, 2005). New sources of data, particularly those associated with online interactions, and innovations in data collection tools are presenting new opportunities to achieve this challenge (Ackland, 2009;Szell and Thurner, 2010). However, others have argued that a more processual orientation to network studies is required (Purchase, Lowe and Ellis, 2010).

Issues of privacy and ethics
Despite the personal nature of much of the data collected and presented in the typical network study, surprisingly little attention has been directed towards addressing the associated issues of privacy and research ethics. Indeed, Breiger (2005, pp. 89-90) pulls no punches in contending that the social network field has 'a greater ability to arrive at incisive analyses than to comprehend the conditions for responsible uses of such analyses'. This is clearly problematic, since as Borgatti and Molina (2003, p. 337) argue, 'In addition to all the usual ethical problems that can arise with any kind of inquiry, network analyses, by their very nature, introduce special ethical problems that should be recognized'. For example, in order to construct a network, the researcher must be able to identify the respondent and the individuals to whom the respondents say they are linked. Thus, although anonymity may be provided at the data presentation stage (i.e. within the network graphic), it is not possible during the data collection stage. Furthermore, network visualizations are 'low-level displays that represent the raw data' rather than 'highly digested outputs of analysis' (Borgatti and Molina, 2003, p. 341), and thus, where they are employed, it is often possible for knowledgeable individuals to identify others within the network even where they have been anonymized.
It is common for network studies to ask personal questions, such as 'Who are you friends with at work?' or 'Who do you socialise with outside of work?' However, despite the use of consent forms, most respondents in network studies will not have been involved in such research before and are unlikely to be aware of how they might feel if they are identified through such questions as being 'marginal' or 'unliked' in their group. Furthermore, where the research forms part of a consultancy project, rather than a piece of academic work, respondents may also be unaware of the possible consequences that might result from subsequent management interventions intended to address features revealed by the network study. As Borgatti and Molina (2003, p. 344) state, 'If subordinates do not understand that their answers on the survey could determine their fate, this could be seen as deceptive and constitute an unethical use of network analysis'.
Interestingly, consent can also be a major problem in SNA with regard to non-participants. Since respondents in SNA research reveal details about their relationships and exchanges with others, the non-participation of an individual in a study does not rule out the possibility that data may be collected about them or that they may be included in subsequent analyses or network depictions. There is also increasing use by academics, consultants and managers of electronic sources of 112

S. Conway
'social' data from social networking sites, chat rooms, blogs and email logs, for example. The privacy and consent issues relating to such data sources have received insufficient serious attention. Hoser and Nitschke (2010) contend that it is not enough to assume the free use of social data simply because it resides in the public domain, arguing for the establishment of a code of behaviour that embraces the notion of 'perceived privacy' (Eyenbach and Till, 2001); thus data posted on a social networking site or newsgroup should only be used 'in the context and by the audience he or she intended it for'.

Implications for network researchers and practitioners
For Borgatti and Molina (2003, p. 337), 'the concept of network has become the metaphor for understanding organizations', among both academics and management consultants. It is within this context that we have sought to provide a critique of the robustness of an increasingly popular approach for revealing and mapping social network structure. This critique is not intended to dismiss the potential of SNA for theory-building or management practice, but rather to surface issues that require consideration and, where possible, resolution or mitigation.

Implications for network researchers and further research
We have argued that the seductive nature of network visualizations has distracted attention away from a number of emerging and longstanding issues in SNA. We contend that network researchers need to reflect more on the choices made concerning boundary-setting and data collection techniques, as well as on the potential impact of missing or inaccurate data. After all, as Rogers (1987, p. 298) has noted, 'without good data, network analysis is worthless'. Indeed, there is a pressing need for further research to improve our understanding of the 'patterns and consequences' of missing network data since, as Kossinets (2006, p. 248) argues, 'Although missing data is abundant in empirical [network] studies, little research has been conducted on the possible effects of missing links or nodes on the measurable properties of networks'.
McGrath, Krackhardt and Blythe (2003, p. 46) also raise concerns about our understanding of to the way in which network visualizations are interpreted by users, arguing 'To be sure, we can make more programs that seem to us as researchers/ programmers to make "better" pictures; but we are relatively ignorant of how general human perception interacts with these fancy new features . . .'. Thus, further research is required in relation to understanding how various users of network maps interpret the visualizations with which they are presented.
It was noted earlier that network studies typically under-emphasize the flows through a network and over-emphasize the quantity rather than the 'quality' or 'utility' of network relationships and interactions (Conway, Jones and Steward, 2001). Such a pattern is likely to be reinforced by the use of network surveys or datamining of social media logs. It is thus recommended that researchers adopt a mixed method approach, incorporating both quantitative and qualitative data collection methods.

Implications for consultants and business practitioners
Network researchers typically construct a single map. Yet, as we have indicated, research has highlighted that individual perceptions (i.e. 'cognitive maps') of a social network can vary greatly from such unified visualizations. This conflation can have far-reaching impacts on the organization since, as Kilduff et al. (2008, pp. 25-26) contend, such 'schema use by individuals in their perceptions of social worlds may affect individuals and larger entities . . . [thus] there may be unanticipated consequences not just for the individuals concerned, but also for the collectivity to which they belong'. Consideration might be given to analysing both the 'cognitive maps' of individual network members and the 'aggregated' network maps produced by network analysts. Social networks are also dynamic in nature; their structure is often fleeting and transitory. Thus, in attempting to make 'invisible' social structures 'visible', network visualizations typically focus attention on the network 'as was' (i.e. when the data were collected) rather than 'as is'. Practitioners must be aware of the implications of the time-lag between data collection and managerial intervention.

Data Inputs and Visual Outputs in Social Network Analysis 113
Practitioners must also recognize that as more network audits are undertaken within their organization it is likely that employees might start to refuse to cooperate, or to complete surveys 'strategically', leading to 'a kind of dialectical arms race' where researchers utilize increasingly sophisticated and passive methods of data collection and employees respond in kind via collusion and manipulation of the data (Borgatti and Molina, 2003, p. 345). Openness with employees in relation to the collection and use of network data within organizations might help to prevent this cycle occurring.
Despite the range of ethical concerns outlined above, Borgatti and Molina (2003, p. 342) argue that what ultimately matters is 'who sees the data and what the data will be used for'. Thus, where the data remain anonymized and do not result in potential consequences for respondents, the ethical 'exposure' may be seen to be greatly reduced. However, these conditions are unlikely to be met where the purpose of the study is to identify appropriate managerial interventions to improve organizational or individual 'performance'. It is also worth researchers seriously considering whether personal questions associated with friendship, within both the work and nonwork environments, are appropriate questions to ask when the study has been commissioned by managers of an organization.
Interestingly, for Kadushin (2005, pp. 139, 151) the question of 'who benefits' is crucial, arguing that 'academic researchers always benefit, organizations, society and science may benefit, but individual respondents rarely do'. The implication of this position is that as network researchers we must become much more sensitized to the range of potential repercussions for respondents. In addressing this issue, some have focused on providing a number of concrete suggestions for the further development of research guidelines and processes (Borgatti and Molina, 2005;Klovdahl, 2005). However, Goolsby (2005) is bolder, contending that there is a need for developing 'an ethical imagination' to tackle these prevailing concerns.