The Effect of Governance on Global Software Development: An Empirical Research inTransactive Memory Systems

Christina Manteli, Bart van den Hooff, Hans van Vliet

VU University, de Boelelaan 1081 Amsterdam, The Netherlands

Abstract

Context: The way global software development (GSD) activities are managed impacts knowledge transactions between teammembers. The first is captured in governance decisions, and the latter in a transactive memory system (TMS), a shared cognitivesystem for encoding, storing and retrieving knowledge between members of a group.

Objective: We seek to identify how different governance decisions (such as business strategy, team configuration, task allocation)affect the structure of transactive memory systems as well as the processes developed within those systems.

Method: We use both a quantitative and a qualitative approach. We collect quantitative data through an online survey to identifytransactive memory systems. We analyze transactive memory structures using social network analysis techniques and we builda latent variable model to measure transactive memory processes. We further support and triangulate our results by means ofinterviews, which also help us examine the GSD governance modes of the participating projects. We analyze governance modes,as set of decisions based on three aspects; business strategy, team structure and composition, and task allocation.

Results: Our results suggest that different governance decisions have a different impact on transactive memory systems. Offshoreinsourcing as a business strategy, for instance, creates tightly-connected clusters, which in turn leads to better developed transactivememory processes. We also find that within the composition and structure of GSD teams, there are boundary spanners (formalor informal) who have a better overview of the network’s activities and become central members within their network. An inter-esting mapping between task allocation and the composition of the network core suggests that the way tasks are allocated amongdistributed teams is an indicator of where expertise resides.

Conclusion: We present an analytical method to examine GSD governance decisions and their effect on transactive memorysystems. Our method can be used from both practitioners and researchers as a “cause and effect” tool for improving collaborationof global software teams.

Keywords: global software development, distributed teams, software governance, transactive memory systems, social networkanalysis

1. Introduction

Over the years globalization of software development activ-ities turned into a common practice. Factors such as the coor-dination and synchronization of the activities across locationsand different time zones, the communication and the knowl-edge management between distributed teams became familiaramong scholars and practitioners. And while research in globalsoftware development (GSD) evolves and new practices emerge[1], Herbsleb [2] notes that “there is little reason to expect thatthese factors will diminish”.

The purpose of this paper is not to try to diminish those fac-tors influencing GSD collaboration, but rather identify themand use them as a tool for a “cause and effect” analysis. Partic-ularly, we are interested in investigating how different decisionsthat companies take on how to govern their GSD activities, af-fect knowledge management processes, and more specificallythe development of transactive memory systems (TMS). In thefollowing paragraphs, we elaborate on that purpose.

With the continuous and evolving strategies in global soft-

ware development, there is a turn of interest towards the chal-lenges and the key issues of managing GSD activities [3, 4, 5].As a result, global software development governance turnedinto an emerging field of research, as a subfield of informationtechnology (IT) governance. The purpose of GSD governanceis to identify those aspects that are necessary for an effective co-ordination and collaboration among distributed teams. As Ban-nerman suggests [6], “governance is the infrastructure neededto ensure the satisfaction of direct and indirect stakeholders”.For instance, when engaged in global activities one can decideto create a captive center in a remote location, or to make a“client-supplier” contract with an external partner. Managersalso need to decide on how to structure their teams, and how toallocate tasks among geographically dispersed members. Whatparts of the projects will be outsourced to the remote partners,and which parts will remain in-site? What kind of responsibil-ities will be delegated to the offshore partners and how muchto delegate? These are all questions that frame the governancestructure that a company builds for its global activities.

Furthermore, different working practices, geographic prox-

Preprint submitted to Information and Software Technology April 10, 2014

imity and/or legal barriers between remote offices, influencethe development of transactive memory systems (TMS) [7].Transactive memory is the kind of memory that the team mem-bers develop and which helps them identify “who knows what”within the team [8]. In order for the team members to de-velop such a memory, they have to engage into various trans-actions through which expertise knowledge is created, sharedand stored. As a result, a cognitive system (transactive memorysystem) is created, where members are aware of each others ex-pertise domain and they are able to access it, update it, share itand facilitate its storage.

Coming back to the main purpose of this study, we pursuethe following research question: How do GSD governance de-cisions affect transactive memory systems? We examine fourcase studies, in two multi-national companies, and we identifytheir governance structure. Based on the different governancedecisions that each case study employs, we report on the differ-ences in the development of transactive memory systems. Wepresent those differences as a “cause and effect” analysis, in or-der to explain how different GSD governance modes (a set ofgovernance decisions) affect the collaboration and communica-tion of the distributed teams.

2. Global Software Development Governance

Research on software development governance is rather re-cent, and as Dubinsky et. al note [9] it is also the result of anincreased focus on the human aspects of software development,such as team work and social collaboration. When the softwaredevelopment activities are globally distributed, the need for aclearly defined governance model increases [10]. For instance,the authors in [11] highlight two “moments” of governance inoutsourcing relationships; the moment of the promissory con-tract which is the formal outsourcing contract between remotepartners, and a second moment of the psychological contract.The latter involves the post-contractual relationship manage-ment that describes the social interactions between the partnerssuch as trust, communication and collaboration aspects.

Several other proposals have been made to define the at-tributes of a software governance model for distributed devel-opment projects, and the coordination mechanisms that such amodel should embrace. For instance, Ramasubbu & Balan [12]present a research model on how to create a governance frame-work for distributed software development, focusing on threestages of a project lifecycle; planning, execution and reflection.In another study, Gewald & Helbig [10] suggest a governanceframework for managing outsourcing engagements based on or-ganizational structures, joint processes and relationship man-agement functions.

In an earlier work [7], we elaborated on global software de-velopment governance and we proposed a GSD governancemodel, based on experiences from an empirical case study. Ourresearch suggested that several knowledge management chal-lenges emerge in GSD settings, caused by the decisions com-panies take on how to govern their global activities, on a strate-gic, tactical and operational level. Building upon the results, wesynthesized a multi-site software governance model, based on

three aspects: the business strategy that binds the mutual rela-tionship of the remote offices, the structure and composition ofthe remote teams and the way tasks are allocated across sites.In the present research, we benefit from this model, and we useit to analyze the governance structures of the current case stud-ies.

• Business Strategy. Carmel and Tija [13] note that one ofthe things that companies should not forget when they op-erate in a global environment is the broader strategic goalsand their legal implications. Recently, a taxonomy wasproposed by Smite et. al [14] in order to categorize andmap different GSD startegies. According to that taxon-omy, GSD business strategies are identified based on thelocation (onshore or offshore), the legal entity (insourceor outsource), and the geographical distance (Near/Far,Close/Distant). To date, two main business strategies arecommonly found in GSD: offshore outsourcing and off-shore insourcing [15]. Offshore outsourcing is when theclient-company works with an external partner in a remotelocation (sometimes referred to as the “buy” decision).Offshore insourcing occurs when the company builds acaptive center in a remote location, such as remote of-fices belong to the same legal entity (also referred to asthe “build” decision).

• Team Structure & Composition. Previous research sug-gests that team structure and composition is a criti-cal factor of good performance in software development[16]. Team size, role descriptions and role distributionare among those characteristics in distributed teams thatcan influence team coordination and communication andtherefore team performance [17, 18]. A prominent rolein the configuration of GSD team is that of “boundaryspanners” [19]. This role is described in various termssuch as “brokers”, “bridges”, “liaisons” and “gatekeep-ers”. Boundary spanning is perceived as a good coordina-tion strategy in the management of distributed collabora-tion, facilitating team awareness and knowledge manage-ment [20]. Boundary spanners are also important within anetwork because they cover the structural holes that mightexist between people, or between sub-groups (clusters).Global software collaboration is, by its nature, a situationwhere structural holes may emerge between groups thatare geographically, temporally and cultural distant [19].Chang & Ehrlich [21] conclude that “individuals who aremore central can exert more influence by virtue of be-ing connected with other powerful individuals in the net-work, and have access to more resources than less centralcounterparts”. Finally, Johri [22] recognized two typesof boundary spanners within organizations; the informalboundary spanners, i.e. people that emerge as media-tors between locations, and formal boundary spanners, i.e.people designated to play that role.

• Task Allocation. Several criteria exist on how to distributework across sites, such as based on the area of expertise,on software architecture and design, or even based on the

2

development steps from requirements elicitation to main-tenance [23]. For instance, a common practice for com-panies that just started outsourcing to a new external part-ner is to delegate in the beginning only small parts of thedevelopment work. This tactic is sometimes referred toalso as “throwing requirements over the fence”. As the re-lationship between distributed offices matures, the clientmight decide to share more responsibilities with the off-shore teams, and work closely together in aspects such asrequirements engineering, architectural decision-makingand system design. Regardless of the way tasks are eventu-ally allocated, work dispersion and the interdependenciesbetween the distributed teams influence the coordinationand administration of the remote offices [24]. For instance,Mockus and Weiss [25] argue that when assigning a task toa remote site, it is important to consider that the allocatedtask matches the capacities of that location. Another studyshows how the increased number of dependencies betweenremote members escalate task complexity and the amountof effort required to share and access knowledge [26].

3. Transactive Memory Systems

A transactive memory system is a shared cognitive system forencoding, storing, and retrieving knowledge between membersof a group. Within TMS, the group members have a collec-tive awareness of each others’ specialized knowledge domainwhich helps them identify and locate where expertise resideswithin the group [27]. Another characterization of a TMS is asa shared understanding of “who knows what” within the group.

Recent studies in global software development turned theirattention to the importance of TMS in the collaboration andcommunication of the distributed teams. One of the main con-cerns in GSD that steered the attention towards TMS is the lackof knowledge that people have about the expertise domain oftheir remote colleagues. Faraj & Sproull [28] propose a modelto test expertise coordination and their empirical results showedthat factors such as locating and sharing expertise knowledgesignificantly improves team performance. In another study[29], the authors identified a positive relationship between dif-ferent dimensions of TMS (such as knowledge differentiation,knowledge location, knowledge credibility) and the communi-cation quality, as well as the performance of the project mem-bers. Coordination of geographically dispersed teams is alsoinfluenced by how well people are aware of each others’ exper-tise [30].

According to Wegner et al. [31], transactive memory hastwo components: an organized store of knowledge (TM struc-ture) and a set of transactions/processes (TM processes) relatedto knowledge management processes of encoding, storing, andretrieving. The TM structure is a knowledge representationof members’ unique and shared knowledge (including mem-bers shared understanding of who knows what). TM processesare the mechanisms by which the group coordinates members’learning and retrieval of knowledge, so that the knowledge canbe applied to group tasks.

In this paper, we examine both the structure of transactivememory systems and the processes involved. The followingsection explains how we measure TM structure and TM pro-cesses accordingly.

3.1. TM Structure and Social NetworksIn order to identify and examine the structure of transactive

memory systems, we adopt a social network approach. So-cial network theory has proven beneficial in analyzing trans-active memory systems. For instance, Yuan et al. [32] studythe social networks of 18 organizational teams and proposedthat within a knowledge sharing network, the strength of tiesbetween the members of the network can influence expertiseawareness. Their findings suggest that although people in con-temporary organizations can learn about each other’s areas ofexpertise through for instance direct interpersonal communica-tions or expertise directories, it is through communication tiesthat employees can gain actual access to diverse expertise. Inanother study, Yuan et al. [33] show that in distributed teams,homophily in social characteristics is an essential factor fordriving the formation of network ties. Additionally, Borgattiet al. [34] developed a model in which the probability of some-one in a network seeking information from another person isa function of (1) knowing what that person knows, (2) valuingwhat that person knows, (3) timely access to that person and (4)the perceived cost associated with accessing that person.

Hence, we use social network measurements and techniquesto analyze transactive memory structures. We do so based onthree fundamental measurements [35]:

1. Clustering, which looks into whether actors of a networktend to form subgroups (clusters) and how strong thoseclusters are.

2. Centrality measures seek to identify team members thatoccupy central positions, and therefore they have an influ-ential role within the structure.

3. Core-periphery structure identifies those actors that aredensely connected and form the core of the network, aswell as those actors that are loosely connected with therest of the network and are part of the periphery.

3.2. TM processes and a latent variable modelThere have been several proposals on how to measure trans-

active memory processes [36, 37, 38]. Lewis & Herndon [39]suggest that transactive memory can be either assessed directlyor indirectly as a latent variable. Direct measures allow us todraw conclusions about TM processes as a whole, while indi-rect measures are useful in settings where TM processes cannotbe clearly assessed. The choice on how to measure TM pro-cesses lies mainly on the case study at hand [39].

In this study, we are interested in examining the effects thatdifferent global software governance decisions have on transac-tive memory systems. In the field of global software develop-ment, transactive memory processes are not clearly defined andhence, we adopt an indirect measure of TMS.

Furthermore, according to Lewis [37] a measure of TMSmust meet two basic criteria. The first is to be theoretically

3

Table 1: TMS OperationalizationDirectory Updating Information Allocation Retrieval Coordination

ExpertiseAwareness

Q1. I am aware of this per-son’s area of expertise.

Q2. I am aware of the knowl-edge that this person needs forhis/her job.

Accessibility Q4. It is easy for me to accessthis person.

Credibility Q5. The information I receivefrom this person is importantfor my job.

CommunicationFrequency

Q3. How often do you transferknowledge to this person?

Q6. How often do you requestknowledge from this person?

consistent with Wegners conceptualization of TMS that reflectsthe cooperative processes and transactions of memory used ina group [8]. The second criterion is to align the measure-ments with the settings of the field we are interested in study-ing. Following these suggestions, we consider previous studiesand conceptualizations of the theory around transactive mem-ory systems and we create a latent variable model to measureTMS in global software development. We choose to create aTMS model, tailored to the concepts of GSD, rather than usean existing one such as the model proposed by Lewis [37],which provides general manifestations of transactive memorysystems. The rest of this section describes how we operational-ized TMS theory and how we used well-established concepts tocreate a latent variable model for TMS processes.

3.2.1. TM operationalizationTable 1 presents the operationalization model that we used

in order to collect the data and express transactive memory asa latent variable. As suggested by Lewis [37], the main driversof our proposed model are the three basic processes of a trans-active memory system: directory updating, information alloca-tion and retrieval coordination [8]. Encoding the information ina memory system implies that group members keep their mem-ory directories updated with regard to what others in the groupare likely to know, or otherwise what is the others’ expertise do-main. Storing the information in a transactive memory systementails the process of allocating knowledge to the person whoseexpertise will facilitate its storage. Additionally, in a knowl-edge network information is retrieved based on knowledge ofthe relative expertise of the individuals.

Previous studies propose that specialization, coordinationand credibility are aspects that can reflect the cognitive pro-cesses within a transactive memory system [28, 37, 40]. Trans-active memory systems exist when members of a group havedifferentiated and specialized knowledge domains, they areaware of the expertise of each other (Expertise awareness),and they can rely on each other for this specialized knowledge(Credibility) [36]. Additionally, in order to share that special-ized knowledge, team members need to develop coordinationactivities that will allow the knowledge to be shared and trans-ferred easily across the network (Accessibility).

Finally, we incorporate into our proposed TMS model, theidea of communication frequency as an important aspect that

needs to be assessed if we want to reflect the existence of TMSwithin distributed teams [34]. Especially in global softwaredevelopment, communication frequency is often regarded as abarrier in the collaboration and knowledge transfer between re-mote team members [41]. In settings where people collaborateacross different geographical locations, it becomes easier fora collective memory system to be built and maintained whencommunication frequency is high.

Summarizing the theoretical foundations of the research, Fig-ure 1 illustrates our conceptual model. Particularly, we explorehow multi-site governance decisions based on business strategy,team structure and composition, and task allocation, impact TMstructure and TM processes.

Figure 1: Conceptual Model

GSD Governance

based on Multi-site Software Governance model

Transactive Memory Systems

TM structure based on social network approach.TM processes based on a latent variable model.

effect

4. Project Overview

The research was conducted in two multi-national compa-nies. The first company, which we call Eco, is developingprinting systems on a global scale. Eco is headquartered inthe Netherlands, with offices in more than 100 countries andover 20,000 employees. Research and development depart-ments work on hardware as well as software innovations, andthey are distributed across nine different countries.

The second company, which we call Ricon, is active in thetelecommunications field in more than 180 countries. Riconemploys over 100,000 people worldwide, offering network plat-forms, telecom services and multimedia solutions for mobilecommunications. During our case study, we worked with theoffices located in the Netherlands, which are involved in hard-ware logistics, as well as in research and development of multi-media solutions.

In the next sections, we elaborate on the projects that partic-ipated in this research. In total, four GSD projects collaboratedwith us, two from Eco and two from Ricon. We are setting the

4

Table 2: Global Software Governance modesBusiness Strategy Team Structure & Composition Task Allocation

Governance mode 1:Eco BU1

Offshore insource. Past ex-perience based on offshoreoutsource.

Boundary spanners are emergent.Offshore teams are hierarchicallystructured.

Close Collaboration: architectural de-cisions and development tasks areshared. System testing and integrationremain in-site.

Governance mode 2:Eco BU2

Offshore outsource Boundary spanners are emergent.Offshore teams are hierarchicallystructured.

Strict separation: part of the develop-ment is outsourced. The rest of the ac-tivities remain in-site.

Governance mode 3:Ricon PR1

Offshore outsource. Boundary Spanners are designated Strict separation: development tasksare entirely outsourced. The rest of theactivities remain in-site.

Governance mode 4:Ricon PR2

Offshore outsource. Pastexperience with offshore in-source.

Boundary spanners are designated. Close collaboration: development tasksare entirely outsourced. The rest of theactivities are shared.

scene of each project, and we identify the governance mode (theset of governance decisions) selected by each case.

Table 2 summarizes the governance modes that we have iden-tified, based on the business strategy, team structure and com-position, and task allocation.

4.1. Setting the scene: Eco

Our research focuses on two business units of Eco company(we refer to them as BU1 and BU2), which specialize in high-end printers. The development of the software used in bothbusiness units is distributed between the main site (site NL)and two remote sites: site A and site B. Figure 2 illustratesthe distributed collaboration between the locations involved inBU1 and BU2. The software includes units such as acceptingrequests, controlling print jobs, rendering images, controllingdevices (e.g. scanners) as well as local and remote user inter-faces.

Figure 2: Eco’s remote collaboration

BU1

site NL

site A

site B

BU2

site NL

site B

Within BU1, site NL and site A work closely together and,from a legal perspective, they are part of the same research anddevelopment department of Eco. Following the taxonomy in[14] the defined business strategy is offshore insource, withnear geographic distance and small temporal distance. Con-sequently, the remote offices are allowed to freely exchangeknowledge and the distributed colleagues collaborate with eachother without any legal barriers limiting their communication.Additionally, the collaboration between site NL and site A is along-term collaboration and the two sites have worked togetherfor many years, as they moved from a client-provider relation-ship to a legal merge.

On the other hand, site B is a different legal entity from siteNL and the collaboration and communication between remoteoffices is rather restricted. Accorind to the taxonomy in [14] thebusiness strategy between site NL and site B can be defined asan offshore outsource strategy where geographic distance is farbut temporal distance is small. Members from site NL are al-lowed to share information with members from site B that refersonly to the tasks they are working on together. General infor-mation and knowledge about other parts of the system can’t beshared with team members of site B.

In a typical team structure, a team leader heads the projectand is responsible for its planning, realization, and success-ful completion. Teams also comprise architects who trans-form the high-level specifications into detailed technical fea-tures and pass them to software engineers who implement thecode. There are also system integrators, responsible for inte-grating the different software units. The communication be-tween distributed teams is not strictly regulated by designatedpeople. Team members are encouraged to proactively shareknowledge and initiate communication with their remote col-leagues. Furthermore, in site NL there is a flat organization andstructure of the teams, without strict reporting lines betweenthe architects, software engineers and testers. Team structuretogether with organization processes differ in the other two re-mote locations, site A and site B. The different tasks and re-sponsibilities between the roles is clearly defined, documentsand issues must be officially accepted before being processed.

Finally the collaboration between remote sites in BU2 in-volves the strict separation of tasks and activities among remote(and co-located) members. The team leader of the specific soft-ware unit allocates development tasks to software engineers.Several criteria are applied in deciding which tasks should besent to site B such as previous experience, i.e., whether a soft-ware engineer has worked on a certain task before, or the com-plexity of the task. Finally, only a small part of the projectis developed in site B, while the majority of the project re-mains in site NL. Within BU1, tasks and activities are allocatedacross teams and across locations in a more interconnected way,where all members work interdependently for the completion ofa project. In the beginning of each release cycle, team leaders

5

and architects from all sites create a plan for the next releaseand discuss the requirements to be implemented. At the end ofeach release, the work developed and tested from the remotesites is shipped back to the NL site for integration and systemand product testing.

4.2. Setting the scene: Ricon

At Ricon’s site, we examined two projects that we refer toas PR1 and PR2. PR1 focuses on the development of a com-munication platform, and it is a collaboration project betweenthe Netherlands (site NL) and two development teams from aremote location, site C. The second project, PR2 involves thedevelopment of various applications that run on the platformdeveloped by PR1. Furthermore, PR2 project concerns the dis-tributed collaboration between site NL and two developmentteams in a remote location, site D (Figure 3). Both site C andsite D are officially external partners, i.e. they are independentsoftware companies with which Ricon has contractual agree-ments. Following the taxonomy in [14] we define the businessstrategy with sites C and D as an offshore outsource strategy,where geographic distance is far. Additionally, temporal dis-tance is small for site C, and large for site D.

Figure 3: Ricon’s remote collaboration

PR1

site NL

site C

PR2

site NL

site D

The decisions for initiating a distributed collaboration withsite C and site D differ, and it is important to take them intoconsideration for the rest of our analysis. Particularly, the col-laboration with site C started based on cost saving decisions,where Ricon was looking for a software development providerin a low-cost location. On the other hand, Ricon had a devel-opment center (captive center) in site D, which closed down.After that, the company in site D re-opened as an independentsoftware vendor, employing the majority of the people workingthere before. The teams of site NL initiated a distributed collab-oration with the new company on a client-provider agreement.The collaboration has an important advantage for the teams insite NL as they work with teams from site D that have alreadya domain-specific knowledge on telecommunications and morespecifically on Ricon’s projects.

Both projects follow the SCRUM methodology and thereforethey employ roles such as scrum masters and product owners.Apart from scrum masters, a development team also involvessoftware developers, testers, and architects. In line with theSCRUM methodology, Ricon keeps small development teamsof five to seven people. Additionally, both projects have ded-icated teams of testers responsible for testing activities at thesystem-level, as well as for system integration. Within Ricon’sprojects we also observe the existence of pre-defined teams and

members responsible for bridging the communication across lo-cations. Particularly, in PR1 there is a dedicated team in siteNL, called team-in-between, which is responsible for the coor-dination of the activities with site C. Team-in-between includesarchitects and testers. Similarly, in PR2 there are specific mem-bers (architects and product owners) from site NL, who are re-sponsible for the coordination of the activities with site D.

To regulate their collaboration with the external partners,Ricon employs what they call internally “steering strategies”.These steering strategies depend on the strategic importanceand/or the complexity of the development work. Managers de-cide how much of the responsibilities to delegate to the externalpartners, ranging from having a team executing only softwaredevelopment tasks, to a self-sustaining and self-managing team.During our case study, within PR1, site C responsibilities arelimited to development activities, with a scrum master in everyteam to coordinate their work. All other decisions and tasks(architecture, system testing, integration) remain in site NL. Onthe other hand, within PR2, site D has more responsibilities in-cluding participating in software architecture decisions and re-quirements engineering processes. At the same time, site NL isalso training members from site D for system testing and inte-gration.

5. Data Collection & Analysis

To analyze the projects and answer our main research ques-tion, we use both quantitative and qualitative data. Surveys arethe most common method for collecting social network data[42, 43]. Additionally, the purpose of our study is to collect dataregarding transactive memory processes, which refer to the cog-nitive memory of individuals. Therefore, we choose to collectsocial network data using an online survey. We distributed thesurvey to all project members of Eco and Ricon. The question-naire invited the respondents to identify those with whom theymainly communicate, in a “free recall” format as described by[43, pg.45]. Additionally, we asked the respondents to value therelationship (on a 5-point Likert scale) with each one of theirconnections, based on the questions Q1-Q6 presented in Table1. The end result is a set of data that represents who communi-cates with whom, and a set of six valued relationships for everyconnection. Overall, BU1 involves 101 people and BU2 con-sists of 70 people. Additionally, we collected data from 54 and48 people from Ricon, working on PR1 and PR2 respectively.

As we previously discussed, TMS is measured as a latentvariable characterizing the link between two people in the net-work (respondent i - connection j), based on six items, i.e. thesix questions of the survey. We performed a reliability analysis,to estimate the degree to which the selected items measure theunderlying construct of transactive memory. A proper analysisfor assessing the reliability of scales is Cronbach’s alpha coeffi-cient [44]. Cronbach’s alpha measures the internal consistency,which is the extent to which all items in a scale can reliablyexpress a latent variable. Cronbach’s alpha coefficient for ourdata set is calculated at .708. This is an acceptable value of in-ternal consistency and hence, we can conclude that our model

6

can be reliably used to measure transactive memory as a latentvariable.

Collecting social network data allows us to express TMSboth as a dyadic relationship between two people, as well asan individual characteristic of each one of the members withina team. In section 3.2, we explained that our indirect way ofmeasuring TMS is guided by the three basic transactions of di-rectory updating (DU), information allocation (IA) and retrievalcoordination (RC). Directory updating is constructed using oneitem (Q1), information allocation is constructed using items Q2,Q3 and finally retrieval coordination is constructed using itemsQ4, Q5, and Q6. Therefore, we calculate TMS at the dyadiclevel according to Eq.(1).

T MS dyadic(i, j) =Q1 + µ(Q2 + Q3) + µ(Q4,Q5,Q6)

3(1)

Subsequently, we need a TMS score that can be attributedto every person in the network, and which will give us a valuefor the transactive memory at the individual level. We calcu-late the individual TMS based on the total scores of T MS dyadic

that an individual has with his direct connections, divided bythe total amount of people in the network, N. Using N as thedenominator in Eq.(2), we take into consideration the unequalamount of direct connections that someone has, by weightinghigher those with more connections and rating lower those withless connections.

T MS individual(i) =

∑ ji=1 T MS dyadic(i, j)

N(2)

Moreover, the data collected from the survey were furthersupported and triangulated by means of interviews with teammembers from both companies. We interviewed members fromall locations in every project involved. We chose a variety ofroles in order to have an accurate representation of the situationin each project. Such roles include software engineers, testers,integrators, architects, and various management roles like teamleaders, project managers, scrum masters, product owners etc.In Table 3, we present details about the number of interviewsand the participants from each project.

Table 3: Number of intervieweesSite NL Site A Site B Site C Site D Total

BU1 (Eco) 10 5 5 - - 20BU2 (Eco) 10 - 5 - - 15

PR1 (Ricon) 5 - - 5 - 10PR2 (Ricon) 3 - - - 4 7

During the interviews, we used a protocol with open-endedquestions to guide the discussions. The protocol had three mainareas of discussion: (a) background information about the inter-viewee, (b) questions about the governance of the project, and(c) questions about the multi-site communication and collabo-ration. The first part of the discussion included questions suchas “What is your role?”, “In which project are you working”and they served as an introductory session.

We used the data from the second part of the interviews todraw the governance modes followed in each project. Some ex-emplary questions included in that part of the discussion are:“How many locations are involved in your project?”, “What isthe contractual relationship between sites?”, “How are tasksallocated between the remote teams?”. The participants ex-plained us whether they have an offshore insource, or offshoreoutsource relationship with the remote offices and how thesecollaborations started. We also discussed with them how tasksare allocated across sites (who does what) and how the teamsare structured and organized in each location. The collectiveresponses helped us draw the different governance modes foreach project, presented in Table 2.

Finally, we were able to understand how people evaluate thecommunication and collaboration with their remote colleagues.The discussions on multi-site collaboration and communicationhelped us understand for instance, how people evaluate commu-nication with their remote colleagues, in terms of how difficultor easy it is for them to share knowledge, be aware of who is theexpert, or collaborate effectively. Eventually, these interviewsprovide further support to the quantitative analysis of transac-tive memory systems. In the next sections, we include quotesfrom the interviews to rationalize the results of our TMS anal-ysis and emphasize our conclusions.

6. Clustering

We begin the analysis with plotting the sociographs of allfour cases, from both Eco and Ricon (Figures 4 and 5). Wealso applied a color-based graphical convention to help withthe interpretation of the sociographs; every color refers to a lo-cation, with orange denoting people from site NL, blue is siteA, black is site B, yellow is site C and finally green refers tosite D. For the structural representation and analysis, we usedUCINET [45] a software package for the analysis of social net-work data, and NetDraw [46] for the graphical representations.The graphical representations (referred to as sociographs fromnow on) were drawn using the spring embedded layout algo-rithm. Based on this algorithm, nodes of the network that havesmall path lengths between them, attract each other and hencethey are closer in the graph [47]. The spring embedded layoutallows us to get a fair overview of the network and the locationof the nodes, by representing which nodes are closer and whichnodes are further apart from each other.

Examining the sociographs, we first notice that people, fromall networks (BU1, BU2, PR1 and PR2), form subgroups (clus-ters). It is also apparent that clustering is based on the geo-graphic location of the team members. For instance, we see thatpeople from site NL are closer to each other, forming their owncluster, while people from sites B, C and D have their own sep-arate clusters. BU1 presents a slightly different picture, wherepeople from site NL and site A are well interconnected, show-ing a weaker tendency to a location-driven separation.

Location-driven clustering is a common structure in globalsoftware development settings, for example in [21, 48, 49].However, our analysis brings up a new observation on the con-nectedness of these clusters. For instance, we notice that in

7

Figure 4: Eco Sociographs Overview

BU1 BU2

Figure 5: Ricon Sociographs Overview

PR1 PR2

BU2, site B has only few connections with site NL, while inBU1, site B is tightly connected with the rest of the network(Figure 4). Similarly, looking at Ricon’s sociographs, we noticethat the clusters in PR2 are better connected, while the clustersin PR1 have fewer links (Figure 5). Based on the interviews,we find that business strategy is a common denominator affect-ing clusters’ connectedness. The remote locations in BU1 wereworking together in the past, in a client-provider contract, andnow they merged into the same legal entity. The remote loca-tions in PR2 were part of the same legal entity, and now theywork in a client-provider relationship. Both cases provide thedistributed teams with years of experience working together,as well as gaining specific-domain knowledge (on printing sys-tems for Eco, on telecommunication systems for Ricon). Even-tually, the business strategy in both cases allow the teams tofreely share knowledge and create effective collaboration pat-terns:

“I have worked with site B for quite some years now, andthey are beginning to transform from a third party that executeswhat we want, into a partner and that makes the system alsobetter.” (System integrator, site NL, BU1)

On the other hand, in BU2 and PR1 people from site NLare engaged into a relationship with external, remote providers,from which they “buy” services. As a result, there are legaldifferences between the remote offices that impede informationallocation and retrieval coordination transactions. External ser-vice providers do not have direct access to all the knowledgestored in site NL, and they rely on the information that theircolleagues selectively send them. Few exemplary quotes:

“If there are problems in sharing information, it is due to

the legal issues. They couldn’t send us, because they are notallowed to. [...] in theory, we have access to all kind of in-formation that they have access to. I say in theory, because inpractice we are different legal entities and there are intellec-tual property issues, so they cannot disclose everything and thiscreates annoyances” (Team leader, site B, BU1)

“We can have only access to that code which we are develop-ing. But if you want to go into details about the platform, thenwe don’t have authorization to access that code or the docu-mentation. If we are facing an issue in that part, then we haveto raise a ticket and they will take it to the respective team.”(Scrum Master, site C, PR1)

Additionally, our results suggest that task allocation can alsoaffect clusters’ connectedness. Specifically, we learn that inBU1 and PR2, remote teams closely collaborate with eachother, taking architectural decisions together, and commonlydefining requirements. This can be the result of an offshoreinsource relationship, arguing that since there are no legal barri-ers and knowledge sharing constraints, teams are free to closelycollaborate with each other. But close collaboration can alsooccur in a offshore outsource relationship, as we see in BU1between site NL and site B. Eventually, a close collaborationincluding shared decision making among distributed teams, andthe delegation of responsibilities to the remote sites create fre-quent interactions between the distributed teams and promoteteam awareness:

“They are more mature, and we also give them more on ar-chitecture. But we work with them, let them make a proposaland then we can look at it, with our architects here, and we de-cide together with them, like ok this should be the approach wehave to take. So they are more involved in the whole processand they really understand what we want from them.” (Productowner, site NL, PR2)

Table 4: TMS dyadic - Descriptive statisticsCompany Project Dyadic relationship Mean SD

EcoBU1 Remote colleagues 3.62 0.63

Co-located colleagues 3.77 0.59

BU2 remote colleagues 3.32 0.52co-located colleagues 3.72 0.41

RiconPR1 remote colleagues 2.55 0.86

co-located colleagues 2.83 0.88

PR2 remote colleagues 2.69 1.02co-located colleagues 2.90 0.99

To further support our observations, we examine transac-tive memory systems from a process-driven approach, usingthe latent variable model described earlier. In order to com-pare TM processes between remote colleagues, we use theT MS dyadic scores, and we distinguish between pairs wheremembers work from different locations. Table 4 presents thedescriptive statistics of all projects. We note that on aver-age co-located colleagues have higher scores than remote col-leagues. When we look at these results independently for ev-ery project, they do not come as a surprise. Nevertheless, ourpurpose here is to analyse TMS scores and examine whetherthat difference is significant, and whether our previous obser-vations on clusters’ connectedness influence that difference.

8

Therefore, we compare the T MS dyadic scores of the remote col-leagues that belong to networks with tightly-connected clus-ters (BU1, PR2), with the scores of the remote colleaguesthat belong to networks with loosely-connected clusters (BU2,PR1). The results support our expectations, suggesting thatTMS processes are better developed between remote colleaguesin BU1 and PR2 (M = 3.29, S D = .90; t(198) = 2.50, p < .05)than between the remote colleagues of BU2 and PR1(M = 2.90, S D = .80; p > .05).

Summarizing our results, we observe that an offshore in-source collaboration eliminates legal barriers between remoteteams and hence, facilitates accessibility and expertise aware-ness. Additionally, when distributed locations closely collab-orate with each other, e.g. taking architectural decisions to-gether, and commonly defining requirements, communicationfrequency and credibility are enforced. Such governance de-cisions have eventually a positive effect in transactive memorysystems; location-driven clusters are tightly connected and themembers develop better transactive memory processes.

7. Centrality & Boundary Spanners

The second step in our analysis is to look at the centralitymeasures of the networks. The centrality measures attempt todescribe the position of the actors in the network and identifythe most prominent, the most central ones. Here we discussthree of the most important centrality measures: degree cen-trality, betweenness centrality and closeness centrality.

Degree centrality is the number of direct connections that anactor has. The idea behind degree centrality measure is that anactor who has many direct connections is also well connectedand therefore that actor can be a central point within the net-work. In terms of communication and knowledge transfer, ac-tors with high degree centrality are considered focal points ofcommunication and major channels of information. Therefore,the degree centrality of an actor is an important index of theactor’s potential communication activity and ability [50].

Betweenness centrality measures how likely an actor is to bethe most direct route between two other actors in the network.Particularly, betweenness centrality is the extent to which anactor lies between two others, acting as a broker or a bridge,managing and mediating information [20]. Additionally, actorswith high betweenness centrality usually seem to have controlover the information that flows between others in the network[51].

Closeness centrality is a measure of how fast someone canreach everyone else in the network. The idea behind closenesscentrality is that an actor is central if he can quickly reach andinteract with all others in the network [43]. Actors with theshortest paths to all others have a good overview, a good in-sight, of what happens in the network. Another study [50] sug-gests that actors with high closeness centrality can be very pro-ductive in communicating information across the network andwhen they are engaged in problem solving they can provide ef-ficient solutions.

In order to identify the most central members of the networksfrom Eco and Ricon, we look at the people that score higher in

all centrality measures (we use the top ten scores from degree,closeness and betweenness results). Next, we annotate thosewho score higher in all centrality measures for every project.In Figure 6 we illustrate the most central members with anenlarged-size node. We also indicate the area within the net-work where the central people are positioned, by drawing a redcircle around them.

Figure 6: Boundary Spanners

BU1 BU2

PR1 PR2

Looking at the centrality results, we observe that the mostcentral members within the network are those that bridge thelocation-driven clusters. In other words, it appears that the mostcentral people within each network are the members that act asboundary spanners between locations. As it was recognized ina previous study, boundary spanning involves activities such asbuilding sustainable relationships, resolving conflicts, negotiat-ing and motivating [52]. Inevitably, boundary spanners becomecentral members within their networks.

As we have described in section 2, Eco does not have des-ignated boundary spanners. The results of the centrality scoresreveal that the most central people in both BU1 and BU2 are theteam leaders, and consequently the boundary spanners betweenlocations. From the interviews, we learn that due to the hier-archical structure of the offshore teams, the role of team leaderemerges as the “interface” between remote locations:

“They are surprised when I’m asking them something with-out involving their team leader. So they are not used to it. ‘Do Ihave to answer this? ...my coach hasn’t agreed on that. shouldI spend time on this?’ ” (Software Engineer, site NL, BU2)

On the other hand, Ricon has designated teams and members,which formally act as boundary spanners between locations:

“We have the product owner and the software architect asmain contacts in site NL. And almost every request or everyquestion should go through them.” (Architect, site D, PR2)

Whether boundary spanners are formal or informal, they re-quire a substantial amount of real-time interactions and a tol-erance for frequent interruptions in order to bring remote col-

9

leagues together, create new relationships and share knowledge[53, 54, 22]. Hence, boundary spanners develop a system-generic knowledge that allows them to have a general overviewof the status of each project and the tasks each team is work-ing on [20]. Based on these suggestions, we would expect thatmembers who are more central in the network, would also havebetter transactive memory systems developed. If someone hasmany direct connections, he is close to many others, and he isin between the communication of many others, then that personis also better aware of the knowledge created, shared and storedacross the network.

In order to assess the levels of TMS for these roles, we com-pare the results of the centrality measures with the scores inT MS individual. The results support our expectations, indicatingthat there is a positive, significant correlation between central-ity measures and TMS processes (Table 5). Note that the scoresof closeness centrality are reverse. The actual scores measure“farness”, hence people with high scores are far from all othersin the network, while people with low scores are close to allothers in the network.

Table 5: Correlation results between centrality measures and TMS processesT MS individual mean SD

Degree .763** 6.37 4.45Closeness -.534** 209.97 85.43

Betweenness .166** 68.27 95.85**correlations significant at p < .001

Summarizing the results, we observe that within the compo-sition and structure of GSD teams, there are boundary spannerswho have a better overview of the network’s activities and bringremote teams together by facilitating knowledge sharing. Dueto these attributes that boundary spanners have, they becomecentral members within their network and they develop bettertransactive memory processes.

8. Network Core

The last step is to look at the core-periphery structures of thenetworks. The core of the network tends to have a more cen-tral position, which indicates that the actors of the core are notonly very well connected with each other but they are also wellconnected with the rest of the actors in the periphery. Core-periphery structure in Open Source Software projects (OSS) issometimes referred to as the onion structure [55]. The core de-velopers are at the center of the onion, they contribute most ofthe code and manage the design of the project [56]. At the pe-riphery of the onion structure, there are developers who commitless to the project by delivering e.g. only bug fixes, use-casesor just testing new releases. The core of OSS projects is impor-tant because many essential tasks for the software developmentprocesses (such as system architecture, conflict resolution, lead-ership etc.) are performed by team members that belong to thatcore [57].

Borgatti & Everett [35] were the first to propose methods fordetecting core-periphery networks. They developed two mod-

els; a discrete model that defines two classes within the net-work, the core and the periphery. This model distinguishes thenetwork members between those that belong to the core, andthose that belong to the periphery. The second model is a con-tinuous model where every member in the network is assigneda measure of “coreness” indicating how close the actor is to thecore.

For our analysis we use both the discrete and the continuousmodel. We use the discrete model to locate the core membersof every project. The results are summarized in Table 6. Wethen use the continuous model to assign all members a corenessvalue, as an individual attribute characterizing their position ina core-periphery structure.

Table 6: Number of core members per locationSite NL Site A Site B Site C Site D Total

BU1 14 8 - - - 22BU2 14 - - - - 14PR1 10 - - 3 - 13PR2 4 - - - 17 20

Interestingly, the analysis of the core reveals a pattern be-tween core members’ distribution across locations and task al-location. Figure 7 shows that there is a correspondence be-tween the tasks and responsibilities of the offshore locations,and the proportion of the core members in the network, origi-nating from these locations.

Figure 7: Task allocation and core composition

BU1 Core compositionTask Allocation

architecture

system testing [I&V]

development

site NL site A

BU2 Core compositionTask Allocation

architecture

system testing [I&V]

development

site NL site B

PR1 Core compositionTask Allocation

development

architecture

system testing [I&V]

site NL site C

PR2 Core compositionTask Allocation

development

architecture

system testing [I&V]

site NL site D

site%NL

Starting from the first case (BU2) in Figure 7, we notice thatsite B is responsible for only a part of the development tasks and

10

Figure 8: The effect of Governance on Transactive Memory Systems

TRANSACTIVE MEMORY SYSTEMS

Business Strategy: • Offshore insource. Remote locations are part of the same legal entity.• There is past experience working together on a client-provider contract.

Team Structure & Composition:• Boundary Spanners are emergent.• Offshore locations tend to be more hierarchically structured.

Task Allocation:• Close collaboration. Architectural decisions and development tasks are shared. System testing and integration remain at the main site.

Business Strategy: • Offshore outsource. Remote locations are part of different legal entities.

Team Structure & Composition:• Boundary Spanners are emergent.• Offshore locations tend to be more hierarchically structured.

Task Allocation:• Strict Separation. Only part of the development tasks are outsourced. The rest of the activities remain in-site.

Business Strategy: • Offshore outsource. Remote locations are part of different legal entities.

Team Structure & Composition:• Boundary Spanners are designated. There is a "team-in-between" bridging remote colleagues.

Task Allocation:• Strict Separation. The development tasks are outsourced. The rest of the activities remain in-site.

Business Strategy: • Offshore outsource. Remote locations are part of different legal entities.• There is past experience working together under the same company.

Team Structure & Composition:• Boundary Spanners are designated. Assigned team members bridge remote colleagues.

Task Allocation:• Close collaboration. The development tasks are outsourced. Architectural decisions and System testing and integration are shared.

Clustering: • Smaller tendency for geographically-based clusters.• Tightly connected clusters.

Centrality:• Team leaders are the central people in the network.

Coreness:• 80% of the core members are in-site.

Clustering: • High tendency for geographically-based clusters.• Tightly connected clusters.

Centrality:• Designated boundary spanners are the central people in the network.

Coreness:• 80% of the core members are out-site.

Clustering: • High tendency for geographically-based clusters.• Loosely connected clusters.

Centrality:• Team leaders are the central people in the network.

Coreness:• 100% of the core members are in-site.

Clustering: • High tendency for geographically-based clusters.• Loosely connected clusters.

Centrality:• Designated boundary spanners are the central people in the network.

Coreness:• 60% of the core members are in-site.

• Within tightly connected, geographically-based clusters, remote colleagues develop better transactive memory processes, than remote colleagues within loosely connected, geographically-based clusters.• The more central a person is in the network, the better developed his transactive memory processes.• The closer to the core of the network a person is, the better developed his transactive memory processes.

GOVERNANCE MODES

effect effect effect effect

THE GOVERNANCE EFFECT

consequently, the entire core of the network consists of mem-bers from site NL only. In the second case (PR1), developmentis the sole responsibility of the offshore location, a decision thatgives site B a small share (20%) in the network core. The thirdcase (BU1) shares architectural decisions as well as develop-ment tasks with the remote teams, and therefore people fromsite A occupy almost half (40%) of the network core. The lastcase (PR2) shows us the “other side of the coin”, where the ma-jority of the core (80%) consists of people from the offshoresite as most of the tasks and responsibilities are delegated to theexternal partner (site D).

As Cataldo & Herbsleb [58] point out “those individuals onthe core, not only perform a critical communication role, butalso they are the top contributors to the actual development ef-fort. In other words, the individuals at the core are also highlyproductive”. We expect that core members in GSD collabora-tive networks, are better aware of the current activities withinthe system, who works on which part of the project and of thecurrent status of the project. Comparing the coreness scoreswith the T MS individual scores, we find that there is a positive,significant correlation between them (r = .588, p < .001).These results suggest that members closer to the core of thenetwork develop better transactive memory processes.

Summarizing our findings, we observe that deciding how toallocate (delegate and/or share) tasks among distributed teamsis an indicator of where expertise resides (in-site or off-site).Core members are well connected with all others in the net-work, they are actively involved in GSD activities and conse-

quently, they develop better transactive memory processes.

9. Lessons Learned

There are no “one size fits all” solutions in global softwaredevelopment. Every company is different and every person isdifferent. With this in mind, we proposed a way to identify andanalyze the different governance decisions of GSD companies,and observe how such decisions affect transactive memory sys-tems. Our findings are summarized in Figure 8. We presentfour different governance modes, based on decisions in busi-ness strategy, team structure and composition, and task alloca-tion. Subsequently, every governance mode affects transactivememory systems on two aspects: on TM structure, which isexpressed through social network measurements of clustering,centrality and core-periphery, and on TM processes measuredas a latent variable.

From a practical point of view, we propose an analyticalmethod that companies may benefit from and use it as a “causeand effect” tool. For instance, GSD companies can use it toevaluate their business strategy with the offshore teams andwhether it leads to strong or weak ties between remote teams aswell as the influence it might have on their knowledge transac-tions. Likewise, managers can analyze their team structure andcomposition and decide whether to assign formal or informalboundary spanners and who should become central member oftheir network. One might also benefit from the core-periphery

11

methods to determine how to allocate tasks, how much to dele-gate to the offshore sites and what tasks to share with the remoteteams.

From a theoretical perspective, we contribute to the body ofknowledge in GSD research in several aspects. First, we presenta systematic way of analyzing governance structures in GSDand their effect in transactive memory systems. Particularly, wesuggest the use of the GSD governance model to analyze multi-site activities based on three aspects; namely here the businessstrategy, task allocation, and team structure and composition.We also propose a two-fold approach to analyze transactivememory systems; a process and a structural approach. As GSDresearch evolves and the attention to knowledge managementchallenges increases, we believe that our contribution will ben-efit future research.

Second, we point out the importance of cluster connected-ness. As we previously discussed, location-driven clusteringis a common phenomenon in GSD, but little attention has beenpaid so far in the significance of the connectedness of such clus-ters. Through our research, we find that business strategy de-cisions such as pursuing an offshore insource approach, createstightly-connected clusters, which in turn leads to better devel-oped transactive memory systems. Third, we contribute to ex-isting research on the role of boundary spanners in GSD, bycomparing cases between formal and informal boundary span-ners. We find that within the composition and structure of GSDteams, boundary spanners play a central role as they have a bet-ter overview of the network’s activities, bring remote teams to-gether and facilitate knowledge sharing. Finally, we presenta novel mapping between task allocation and core-peripherystructures in GSD networks. We show how from sharing smalltasks with the offshore teams to delegating and sharing moreresponsibilities, affects the composition of the network core.

10. Limitations

The credibility of a study refers to the degree of confidencewe have on its conclusions [59]. Although we have addressedthe validity of our study from the beginning of this research, itis important to point out several aspects that might be a threat.We do so for two reasons; first because we want to establishthe integrity and credibility of our observations and results, andsecond for suggesting future improvements.

To tackle internal validity in our study, we used data triangu-lation. We collected quantitative data from the online survey,and we further supported our observations with qualitative datafrom the interviews. Nevertheless, collecting data using ques-tionnaires might be biased from the respondents individual per-ceptions of reality. We propose that a replication of the methodused in this study by collecting data from other sources suchas documentation or email communication will improve the in-ternal validity of the current study. Furtheremore, the absenceof certain individual attributes of the participants, such as theyears of experience working in the company/project, might alsoinfluence internal validity. For instance, someone who is work-ing for many years in the project might have a better overviewof the project, its members and the knowledge flow within the

network, compared to newly introduced members. Cultural as-pects might also influence people’s responses both during theinterviews and when filling in the online survey. In this researchwe do not take these aspects into account, but we suggest theirconsideration for the future. Additionally, real networks arenot static, but rather dynamic. Relationships between peoplechange and as such, communication networks evolve. In thecurrent study, we examine only a snapshot of the networks. Aninteresting approach for a subsequent study is to compare trans-active memory systems over time and observe their evolution.

While we report on the reliability analysis for the latent vari-able model of TM processes, we recognize that there might bethreats to construct validity. We propose that the theoreticalconcepts and the questions used for the operationalization ofthe model be further re-evaluated, tested, validated and poten-tially improved.

Finally, we need to be cautious on how broadly we interpretand generalize the results. Although we used four different casestudies, our observations might be influenced by particular anddistinctive characteristics of the companies, such as the com-pany culture, trust issues between remote locations and others.Whereas “a theory may never be generalized to a setting whereit has not yet been empirically tested and confirmed” [60], wesuggest that using additional case studies to the research caneliminate the threat to external validity.

11. Conclusion

In this paper, we examined how different governance deci-sions influence transactive memory systems. We analyzed gov-ernance decisions according to business strategy, task allocationand, team structure and composition. Additionally, we useda two-fold approach in analysing transactive memory systems,based on the structure as well as the processes involved in suchsystems. We conclude with the presentation of a “cause andeffect” method that we can use to examine how different gov-ernance decisions (the cause), affect in different ways the struc-ture and the processes of TMS (the effect).

Our results suggest that looking at the connectedness of thedistributed locations (clusters) involved in the GSD activities,we can examine how well (or not so well) remote colleagues areaware of each other’s expertise knowledge. Furthermore, iden-tifying the central members of a GSD collaboration networkwe can locate those members that span the boundaries acrosslocations and bridge expertise knowledge between distributedteams. Finally, we learned that analyzing the composition ofthe core of a GSD network, we can explore where the most ac-tive members of the project are located.

We contribute to the body of knowledge in the field ofglobal software engineering, by proposing a different analysisof multi-site collaborations. We merge the business and gov-ernance aspects of a GSD with the evergrowing challenges inknowledge management and expertise awareness, using as a“glue” social network analysis techniques. Finally, we encour-age future research towards the further enhacement and devel-opment of the use of transactive memory systems and socialnetworks in global software development studies.

12

References

[1] D. Damian, D. Moitra, Global software development: How far have wecome?, Software, IEEE 23 (5) (2006) 17–19.

[2] J. D. Herbsleb, Global software engineering: The future of socio-technical coordination, in: Future of Software Engineering (FOSE), IEEEComputer Society, 2007, pp. 188–198.

[3] S. L. Jarvenpaa, D. E. Leidner, Communication and trust in globalvirtual teams, Journal of Computer-Mediated Communication 3 (4).doi:10.1111/j.1083-6101.1998.tb00080.x.

[4] T. Dingsøyr, R. Conradi, A survey of case studies of the use of knowl-edge management in software engineering, International Journal of Soft-ware Engineering and Knowledge Engineering 12 (04) (2002) 391–414.doi:10.1142/S0218194002000962.

[5] A. Tiwana, An empirical study of the effect of knowl-edge integration on software development performance, In-formation and Software Technology 46 (13) (2004) 899–906.doi:http://dx.doi.org/10.1016/j.infsof.2004.03.006.

[6] P. L. Bannerman, Software development governance: A meta-management perspective, in: Proceedings of the 2009 ICSE Workshopon Software Development Governance, IEEE Computer Society, 2009,pp. 3–8.

[7] C. Manteli, B. van den Hooff, A. Tang, H. van Vliet, The impact of multi-site software governance on knowledge management, in: Global SoftwareEngineering (ICGSE), 2011 6th IEEE International Conference on, IEEE,2011, pp. 40–49.

[8] D. M. Wegner, Transactive memory: A contemporary analysis of thegroup mind, in: Theories of group behavior, Springer, 1987, pp. 185–208.

[9] Y. Dubinsky, S. Ravid, A. Rafaeli, R. Bar-Nahor, Governance mech-anisms in global development environments, in: Global Software En-gineering (ICGSE), 2011 6th IEEE International Conference on, IEEE,2011, pp. 6–14.

[10] H. Gewald, K. Helbig, A governance model for managing outsourc-ing partnerships: a view from practice, in: System Sciences, 2006.HICSS’06. Proceedings of the 39th Annual Hawaii International Con-ference on, Vol. 8, IEEE, 2006, pp. 194c–194c.

[11] S. M. Miranda, C. B. Kavan, Moments of governance in is outsourcing:conceptualizing effects of contracts on value capture and creation, Journalof Information Technology 20 (3) (2005) 152–169.

[12] N. Ramasubbu, R. K. Balan, Towards governance schemes for distributedsoftware development projects, in: Proceedings of the 1st internationalworkshop on Software development governance, ACM, 2008, pp. 11–14.

[13] E. Carmel, P. Tjia, Offshoring information technology: sourcing and out-sourcing to a global workforce, Cambridge University Press, 2005.

[14] D. Smite, C. Wohlin, Z. Galvina, R. Prikladnicki, An empirically basedterminology and taxonomy for global software engineering, EmpiricalSoftware Engineering (2012) 1–49.

[15] M. Robinson, R. Kalakota, Offshore outsourcing: Business models, ROIand best practices, Mivar Press, 2004.

[16] H.-L. Yang, J.-H. Tang, Team structure and team performance in is devel-opment: a social network perspective, Information & Management 41 (3)(2004) 335–349.

[17] M. B. O’Leary, M. Mortensen, Go (con) figure: Subgroups, imbalance,and isolates in geographically dispersed teams, Organization Science21 (1) (2010) 115–131.

[18] A. Kofman, A. Yaeli, T. Klinger, P. Tarr, Roles, rights, and responsibili-ties: Better governance through decision rights automation, in: Proceed-ings of the 2009 ICSE Workshop on Software Development Governance,IEEE Computer Society, 2009, pp. 9–14.

[19] A. E. Milewski, M. Tremaine, F. Kobler, R. Egan, S. Zhang, P. O’Sullivan,Guidelines for effective bridging in global software engineering, SoftwareProcess: Improvement and Practice 13 (6) (2008) 477–492.

[20] A. Boden, G. Avram, Bridging knowledge distribution - the role of knowl-edge brokers in distributed software development teams, in: Proceedingsof the 2009 ICSE Workshop on Cooperative and Human Aspects on Soft-ware Engineering, CHASE ’09, IEEE Computer Society, Washington,DC, USA, 2009, pp. 8–11.

[21] K. T. Chang, K. Ehrlich, Out of sight but not out of mind?: Informalnetworks, communication and media use in global software teams, in:Proceedings of the 2007 conference of the center for advanced studies onCollaborative research, IBM Corp., 2007, pp. 86–97.

[22] A. Johri, Boundary spanning knowledge broker: An emerging role inglobal engineering firms, in: Frontiers in Education Conference, 2008.FIE 2008. 38th Annual, IEEE, 2008, pp. S2E–7.

[23] A. Lamersdorf, J. Munch, D. Rombach, A survey on the state of the prac-tice in distributed software development: Criteria for task allocation, in:Global Software Engineering, 2009. ICGSE 2009. Fourth IEEE Interna-tional Conference on, IEEE, 2009, pp. 41–50.

[24] J. D. Herbsleb, A. Mockus, An empirical study of speed and communica-tion in globally distributed software development, Software Engineering,IEEE Transactions on 29 (6) (2003) 481–494.

[25] A. Mockus, D. M. Weiss, Globalization by chunking: a quantitative ap-proach, Software, IEEE 18 (2) (2001) 30–37.

[26] P. R. Carlile, Transferring, translating, and transforming: An integrativeframework for managing knowledge across boundaries, Organization sci-ence 15 (5) (2004) 555–568.

[27] A. B. Hollingshead, Communication, learning, and retrieval in transac-tive memory systems, Journal of experimental social psychology 34 (5)(1998) 423–442.

[28] S. Faraj, L. Sproull, Coordinating expertise in software developmentteams, Management science 46 (12) (2000) 1554–1568.

[29] X. Chen, X. Li, J. G. Clark, G. B. Dietrich, Knowledge sharing in opensource software project teams: A transactive memory system perspective,International Journal of Information Management.

[30] J. A. Espinosa, S. A. Slaughter, R. E. Kraut, J. D. Herbsleb, Team knowl-edge and coordination in geographically distributed software develop-ment, Journal of Management Information Systems 24 (1) (2007) 135–169.

[31] D. M. Wegner, T. Giuliano, P. T. Hertel, Cognitive interdependencein close relationships, in: Compatible and incompatible relationships,Springer, 1985, pp. 253–276.

[32] Y. C. Yuan, J. Fulk, P. R. Monge, N. Contractor, Expertise directory de-velopment, shared task interdependence, and strength of communicationnetwork ties as multilevel predictors of expertise exchange in transactivememory work groups, Communication Research 37 (1) (2010) 20–47.

[33] Y. C. Yuan, G. Gay, Homophily of network ties and bonding and bridg-ing social capital in computer-mediated distributed teams, Journal ofComputer-Mediated Communication 11 (4) (2006) 1062–1084.

[34] S. P. Borgatti, R. Cross, A relational view of information seeking andlearning in social networks, Management science 49 (4) (2003) 432–445.

[35] S. P. Borgatti, M. G. Everett, Models of core/periphery structures, Socialnetworks 21 (4) (2000) 375–395.

[36] J. R. Austin, Transactive memory in organizational groups: the effects ofcontent, consensus, specialization, and accuracy on group performance.,Journal of Applied Psychology 88 (5) (2003) 866.

[37] K. Lewis, Measuring transactive memory systems in the field: scale de-velopment and validation., Journal of Applied Psychology 88 (4) (2003)587.

[38] Y. Yoo, P. Kanawattanachai, Developments of transactive memory sys-tems and collective mind in virtual teams, International Journal of Orga-nizational Analysis 9 (2) (2001) 187–208.

[39] K. Lewis, B. Herndon, Transactive memory systems: Current issues andfuture research directions, Organization Science 22 (5) (2011) 1254–1265.

[40] R. L. Moreland, L. Myaskovsky, Exploring the performance benefits ofgroup training: Transactive memory or improved communication?, Orga-nizational Behavior and Human Decision Processes 82 (1) (2000) 117–133.

[41] R. Noordeloos, C. Manteli, H. V. Vliet, From rup to scrum in globalsoftware development: A case study, in: Global Software Engineering(ICGSE), 2012 IEEE Seventh International Conference on, IEEE, 2012,pp. 31–40.

[42] M. Newman, Networks: an introduction, Oxford University Press, 2009.[43] S. Wasserman, K. Faust, Social Network Analysis: Methods and Appli-

cations, Vol. 8, Cambridge University Press, New York, USA, 1994.[44] L. J. Cronbach, Coefficient alpha and the internal structure of tests, Psy-

chometrika 16 (3) (1951) 297–334.[45] S. P. Borgatti, M. G. Everett, L. C. Freeman, Ucinet for windows: Soft-

ware for social network analysis.[46] S. Borgatti, Netdraw software for network visualization, Lexington, KY:

Analytic Technologies.[47] R. A. Hanneman, M. Riddle, Introduction to social network methods.

13

[48] M. Cataldo, K. Ehrlich, The impact of communication structure on newproduct development outcomes, in: Proceedings of the 2012 ACM annualconference on Human Factors in Computing Systems, ACM, 2012, pp.3081–3090.

[49] K. Ehrlich, C.-Y. Lin, V. Griffiths-Fisher, Searching for experts in theenterprise: combining text and social network analysis, in: Proceedingsof the 2007 international ACM conference on Supporting group work,ACM, 2007, pp. 117–126.

[50] L. C. Freeman, Centrality in social networks conceptual clarification, So-cial networks 1 (3) (1979) 215–239.

[51] M. E. Newman, A measure of betweenness centrality based on randomwalks, Social networks 27 (1) (2005) 39–54.

[52] P. Williams, The competent boundary spanner, Public administration80 (1) (2002) 103–124.

[53] A. B. Hargadon, Brokering knowledge: Linking learning and innovation,Research in Organizational behavior 24 (2002) 41–85.

[54] N. Levina, Collaborating across boundaries in a global economy: Do or-ganizational boundaries and country contexts matter?

[55] J. Y. Moon, L. Sproull, Essence of distributed work: The case of the linuxkernel, Distributed work (2002) 381–404.

[56] K. Crowston, J. Howison, The social structure of free and open sourcesoftware development, First Monday 10 (2).

[57] K. Crowston, K. Wei, Q. Li, J. Howison, Core and periphery in free/libreand open source software team communications, in: System Sciences,2006. HICSS’06. Proceedings of the 39th Annual Hawaii InternationalConference on, Vol. 6, IEEE, 2006, pp. 118a–118a.

[58] M. Cataldo, J. D. Herbsleb, Communication networks in geographicallydistributed software development, in: Proceedings of the 2008 ACM con-ference on Computer supported cooperative work, ACM, 2008, pp. 579–588.

[59] D. E. Perry, A. A. Porter, L. G. Votta, Empirical studies of software engi-neering: a roadmap, in: Proceedings of the conference on The future ofSoftware engineering, ACM, 2000, pp. 345–355.

[60] A. S. Lee, R. L. Baskerville, Generalizing generalizability in informationsystems research, Information systems research 14 (3) (2003) 221–243.

14