Computers in Human Behavior

Computers in Human Behavior 27 (2011) 490 503 Contents lists available at ScienceDirect Computers in Human Behavior journal homepage: www.elsevier.com/locate/comphumbeh Online collaboration: Collaborative behavior patterns and factors affecting globally distributed team performance Fatma Cemile Serçe a,, Kathleen Swigger b, Ferda Nur Alpaslan c, Robert Brazile b, George Dafoulas d, Victor Lopez e a Department of Information Systems Engineering, Atilim University, Ankara, Turkey b Department of Computer Sciences and Engineering, University of North Texas, Denton, TX, USA c Department of Computer Engineering, Middle East Technical University, Ankara, Turkey d School of Computing Science, Middlesex University, London, UK e Facultad de Ingeniería de Sistemas Computacionales, Universidad Tecnologica de Panama, Panama article info abstract Article history: Available online 20 October 2010 Keywords: Collaborative learning Computer-supported collaborative learning Distributed teams Collaborative behavior Global software development K-means clustering Studying the collaborative behavior of online learning teams and how this behavior is related to communication mode and task type is a complex process. Research about small group learning suggests that a higher percentage of social interactions occur in synchronous rather than asynchronous mode, and that students spend more time in task-oriented interaction in asynchronous discussions than in synchronous mode. This study analyzed the collaborative interaction patterns of global software development learning teams composed of students from Turkey, US, and Panama. Data collected from students chat histories and forum discussions from three global software development projects were collected and compared. Both qualitative and quantitative analysis methods were used to determine the differences between a group s communication patterns in asynchronous versus synchronous communication mode. K-means clustering with the Ward method was used to investigate the patterns of behaviors in distributed teams. The results show that communication patterns are related to communication mode, the nature of the task, and the experience level of the leader. The paper also includes recommendations for building effective online collaborative teams and describes future research possibilities. Ó 2010 Elsevier Ltd. All rights reserved. 1. Introduction Advances in technology as well as changes within many organizations have led to an increased interest in hiring people who can communicate effectively within a distributed environment. Instead of meeting face-to-face with colleagues in offices, distributed workers rely on special communication technologies, such as e-mail, real-time chat tools, teleconference software, and virtual meeting rooms, to establish relationships and do business. Because electronic media often limits the amount of non-verbal exchanges among team members, the success and effectiveness of distributed teams is much more sensitive to a member s skill in using technology to communicate and collaborate. In order to be competitive within the new work environment, students must learn how to develop skills that facilitate communication and collaboration across distances and time. Because of these growing changes within the market place, educators are being pressured to teach students how to become more Corresponding author. Tel.: +90 312 586 83 75; fax: +90 312 586 80 91. E-mail address: fcserce@atilim.edu.tr (F.C. Serçe). proficient at distributed communication and collaboration. The widespread use of computer-supported collaborative learning (CSCL) environments that are supported by learning management systems is allowing educators to explore ways to promote collaborative communication skills among students attending colleges and universities. These collaborative learning platforms offer a variety of tools that facilitate information sharing and communication among participants. Although educators have tended to favor the asynchronous tools such as forums and wikis for collaboration, they have shown an increasing interest in synchronous tools that can be used to enhance a group s social interactions (Lobel, Neubauer, & Swedburg, 2002; Locatis et al., 2003). Factors such as the difficulty in coordinating meeting times, the high cost of quality synchronous communication technology, and tool stability may explain the underutilization of synchronous tools (Burnett, 2003). However, as these synchronous technologies have improved and become more affordable, they have also become more established in distance-learning environments. Thus, there is a growing need for research that can help determine the strengths and weaknesses of different communication modes so that they can be better understood and utilized more effectively. 0747-5632/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.chb.2010.09.017

F.C. Serçe et al. / Computers in Human Behavior 27 (2011) 490 503 491 Global software engineering courses offer researchers a unique opportunity for looking at differences among modes of communication, particularly as they might affect collaboration among team members who are given a specific task to solve. While computer science and IT educators were not early adopters of learning management systems, they have since realized the importance of teaching students how to use different computer-supported tools for group work. Software development courses are designed to teach students how to work in teams on large software projects. More recently, the software development courses have included more realistic projects that require students to work with team members who are located in other countries and/or time zones. Since global software development projects emphasize communication, coordination, and control (Holmstrom, Conchuir, Agerfalk, & Fitzgerald, 2006), students must learn how to complete each of these processes using one or more computer-supported collaborative tools (Gotel, Scharff, & Seng, 2006; Lanubile, Damian, & Oppenheimer, 2003). Student groups enrolled in global software courses learn how to use various online tools to build teams, exchange information, and work on projects together. They also provide an environment for teaching students how to manage groups that have different cultures and live in different time zones. Thus, global software development courses provide an excellent context for investigating the communication modes that contribute to the exercise of different collaborative interaction patterns among learning teams. In order to address this issue, we began a 3-year study to investigate team effectiveness factors among global software development learning teams. One of the important components within the context of global software development is the communication among group members and how it may affect a team s performance. Thus, one of the objectives of this project is to understand how students use new technology to communicate and share ideas, code, and information. Each semester, students from the three participating universities (i.e., Atılım University, Panama Technology University, and University of North Texas) are grouped together and asked to complete a software development project. The classroom projects are intended to mimic the inherent global software development (GSD) characteristics of geographical distance, different cultures, and different time zones. Using various computer-supported collaborative tools, students learn how to communicate with their teammates and coordinate the different software development tasks. Because these interactions are recorded, we are able to examine the different communication activities in an effort to determine which factors lead to better performance. These particular analyses are designed to give us useful insights into the dynamics that may affect distributed teams. They also provide a basis for helping students learn how to either maximize or minimize the various factors that might lead to more successful collaborations. The research presented in this paper is an attempt to determine if different communication modes affect the communication patterns of global software learning teams. The implication for such a study is that the results will help teachers of distributed learning courses (everywhere) provide students with more informed instruction about how to communicate with team members. In order to address these issues, the authors analyzed the computer conferencing transcripts of three different global software development student projects by means of a content classification scheme developed by Curtis and Lawson (2001). We then examined the results of these classifications and compared the patterns of communication behaviors that occurred within each project to determine whether there were any differences among groups based on their use of asynchronous versus synchronous communication tools. The paper begins with a report on the relevant research that was used to guide this study, followed by an overview of the experiment. The paper also includes a description of the coding scheme and the measures that were used to gather data about individuals and teams. Finally, the paper presents the results of our analyses and concludes with a list of recommendations that are meant to improve future work. 2. Previous work 2.1. Asynchronous versus synchronous technologies Educational collaborative technologies are generally defined as the software (and sometimes hardware) that support and enhance group communication and teamwork in academic courses. There are two general types of collaborative technologies: asynchronous and synchronous. Asynchronous collaborative technologies refer to the category of communication tools that allow people to share ideas on their own time. All the participants of a group need not be present for communication to occur; rather, individuals can send/post information to team members whenever they are available. Learning events tend to be independent and unattached to a particular time or space, so learners participate when and if they choose (Rovy & Essex, 2001; Sabau, 2005). On the other hand, synchronous collaborative technologies are communication tools that allow people to collaborate in real time. All participants must be present for the communication, and any learning that occurs is attached to a particular time (Romiszowski & Mason, 2004). The majority of the computer-supported collaborative learning (CSCL) research has been done using asynchronous communication modes. The reason for the widespread use of asynchronous communication tools in educational environments is because these are usually easier to implement and support (Branon & Essex, 2001). Developers of online learning environments often suggest that asynchronous communication may have advantages over synchronous and are the preferred mode of discussion (Bannon, 1995; Dede & Kremer, 1999). For example, Driscoll (1999) claims that asynchronous methods allow students more time for reflection than do synchronous delivery formats. They also tend to be more flexible, allowing all students to respond to a topic (Wang, 2004). There is also some evidence that indicates students prefer asynchronous communication, and that they tend to be more satisfied with their performance when using such tools (Walker & Arnold, 2004). Research on the effectiveness of synchronous communication for collaboration is relatively recent. Previous research about the use of synchronous chat tended to focus on simple case studies that indicated the feasibility of the technology (Locatis et al., 2003). While largely anecdotal, these case studies suggest that synchronous text-based communication contributes to student motivation and specific learning outcomes, and there is some evidence to suggest that students enjoy synchronous chat as a supplement to other learning events (Dickey, 2003; Shotsberger, 2000). More recently, studies have shown that students seem to generate more discussion and ideas using synchronous tools (Koory, 2003). In a more recent survey, Soller, Martinez, Jermann, and Muehlenbrock (2005) characterize the different synchronous/ asynchronous tools in different learning systems according to how they manage the overall collaborative interaction. For example, some systems assimilate the interactions for groups and provide feedback to both students and teachers. Other systems respond to active learning events by providing coaching information to students. Thus, the synchronous or asynchronous tools are viewed with respect to how much feedback they provide to a collaboration event (Soller et al., 2005). While preliminary research suggests that there is a relationship between synchronous and asynchronous online discussion, there is relatively little systematic research on the ways in which

492 F.C. Serçe et al. / Computers in Human Behavior 27 (2011) 490 503 synchronous and asynchronous online discussion can be combined to maximize group collaboration. This research investigates the factors that contribute to the different interaction patterns in different communication modes and how this may be linked to different types of tasks. 2.2. Computer-supported collaborative learning (CSCL) Collaborative learning, teamwork, student-centered learning and students taking responsibility for their own learning are all important objectives for educators in the information age. Both employers and government officials are demanding that students graduate with good interpersonal skills, knowledge of group dynamics, and understanding of work teams. Industry leaders also insist that students be taught how to lead, problem-solve and communicate effectively. Anuradha (1995) argues that if the purpose of instruction is to enhance critical-thinking and problem-solving skills, then students will benefit most from activities that support collaborative learning. The basic principle of collaborative learning is to organize groups of learners to work together to accomplish some shared goal. There have been numerous studies that report on the successful use of CSCL. For example, Cavanaugh (2001) showed that there was no difference in the academic performance of students enrolled in distance education versus those in face-to-face classes. The author also suggests that CSCL can be used to complement, enhance and expand educational options for students. Lou, Abrami and d Apollonia (2001) demonstrated that small group learning with CSCL has more positive effects on a student s cognitive processes and affective outcomes than individual learning. This observation was based on an examination of 486 independent results reported in 122 studies involving 11,317 learners. Hollenbeck (1998) found that distance learners were more likely to take control of their own learning and to perform instructor-like tasks such as discussion and summarizing. Such activities ultimately allow students to feel more autonomous about their learning objectives (Lazakidou & Retalis, 2010). Although the perceived potential of CSCL seems to be at least partially supported by research, there are also studies that suggest there are problems when students work in groups via computers. Communication among members in a group can be both an intentional and an unintentional process. That is, words, phrases, sentences, can be formulated with great detail and purpose such as in a speech or lecture. Much is also, however, communicated unintentionally through non-verbal communications or implication (Nolan, 2006). If these unintentional cues are lost, then members of the team may be conflicted. For example, Hobman, Bordia, Irmer, and Chang (2002) stated that collaborative groups exhibit more process conflicts than face-to-face groups in the early stages of a project. Studies also report that students sometimes have difficulty following the conversation in a chat session, resulting in a condition described as Chat Confusion (Thirunarayanan, 2000). Finally, Fung (2004) found that students enrolled in a distance-learning course, as opposed to a traditional classroom, asked fewer questions and shared fewer ideas. Students enrolled in the distancelearning course reported that a lack of time was the major reason for their limited posts (Branon & Essex, 2001). There have also been several studies that have looked at different factors that may affect team collaboration. Janssen et al.(2009) found that groups who were more familiar with one another were more cohesive because they were able to move more quickly through the early stages of group development (Janssen, Erkens, & Kanselaar, 2007; Janssen, Erkens, Kirschner, & Kanselaar, 2009). Task, group formation, language, and culture have also been found to affect collaboration and performance (Bannon, 1995; Dillenbourg, Baker, Blaye, & O malley, 1996; Freedman & Liu, 1996; Iivonen, Sonnenwald, Parma, & Poole-Kober, 1998; Warschauer, 1997). Research has also shown that practical learning tasks result in deeper processing of information and more affective collaboration (Delaat, Lally, Simons, & Wenger, 2006). Empirical studies have also examined the effects of asynchronous versus synchronous communication modes on CSCL. Although some researchers acknowledge the contribution of synchronous communication technologies to student learning, most argue that asynchronous communication technologies are more effective in supporting collaborative learning among learners (Hiltz, 1998; Ligorio, 2001). Asynchronous communication technologies provide students with time to think about a problem, and the opportunity to discuss solutions in a group independent from time and space (Hiltz, 1998). Because of their flexibility, asynchronous technologies are considered to be essential for creating collaborative and cooperative distance-learning environments (McIsaac & Gunawardena, 1996). Although asynchronous communication technologies continue to be the most common application used in online courses, (Klobas & Haddow, 2000), researchers are now looking at how different modes of communication might be used to increase the amount of social interaction in online learning environment (Hiltz, 1998). Whenever computer-supported collaborations involve student groups from multiple countries or universities, then it is necessary to examine cultural distance and how it impacts the performance of student learning teams (Swigger, Alpaslan, Brazile, Harrington, & Peng, 2005). Cultural values include those traits associated with a specific country or national ethos (Aguinis & Henle, 2003) as well as those connected to a personal community (e.g., school or region) (Fruchter & Townsend, 2003). Cultural diversity can be defined on a number of levels, including national (country), organizational, professional, or team. National culture includes those traits associated with a specific country or national ethos such as language and religion and is often associated with particular geographical boundaries (Hofstede, 1991). Organizational culture encompasses the unit s norms and values, where the unit can range from a city, company, or university (Herbsleb, 2007). These different cultural factors occur on many levels and can coexist, interact with one another, and sometimes collide. While computer-mediated technologies are able to shorten some of cultural distances, they may not be able change students preconceived notions of the other university and the other students (Layzell, Brereton, & French, 2000). Thus, it is necessary to evaluate the impact of culture on the collaboration of student teams whenever multiple universities ore countries are involved. To receive the full benefit of CSCL, a collaborative learner must interact, share information and coordinate actions (Orvis & Lassiter, 2008). As Dillenbourg, Baker, Blaye, and O malley (1996) argue, student teams do not learn merely because there are no two students instead of one, but because those two students perform activities that trigger specific learning mechanisms. The interaction among subjects generates additional activities (explanation, disagreement, and mutual regulation), triggering knowledge elicitation, internalization, and a reduction in cognitive load. In this study, we characterize the different patterns of communication behaviors that trigger online collaboration and examine how different patterns are shaped by communication mode. 2.3. Exploring communication behaviors in CSCL Many of the approaches for analyzing CSCL data have concentrated on low-level quantitative measures such as duration of communication or time of communication. For example, studies claim that there are relationships between participation level and the number of posts to forums (Harasim, Calvert, & Groeneboer, 1996), and the quality of the message content and mean number

F.C. Serçe et al. / Computers in Human Behavior 27 (2011) 490 503 493 of words in a message (Benbunan-Fich, Hiltz, & Turoff, 2003). While such measures give us a rough analysis about the quality of interaction, they provide little information about a message s content (Strijbos, Martens, & Jochems, 2006). Recent research has examined more process-oriented behaviors such as the patterns of communication and classifications along psychological and/or linguistic dimensions. European efforts such as the Kaleidoscope (http://www.noe-kaleidoscope.org/pub/) and TELL projects (http://cosy.ted.unipi.gr/tell/) have published a number of reports about the use of interaction analysis tools to evaluate computer-supported collaborations. These newer interaction analysis tools allow teachers to have a deeper understanding of the learning process (Petropoulou, Lazakidou, Retalis, & Vrasidas, 2007; Xenos, Avouris, Stavrinoudis, & Margaritis, 2009). These content approaches to understanding CSCL usually involve creating a coding scheme that represents all the interesting categories of a particular type of communication such as the rules being displayed in the conversation (Emmert & Barker, 1989), the types of speech (Contractor & Grant, 1996), or the actual meaning of the discussion (Fortuna, Mendes, & Milic-Frayling, 2007). The transcribed discourse is then divided into the smallest units of meaning, and those pieces of text that correspond to the categories of interest are tagged (Emmert & Barker, 1989). For example, Walther (1996) proposes a three-level coding scheme to measure the aspects of interactions in computer-mediated communications, namely impersonal, interpersonal, and hyperpersonal interaction. In another study, Bonk and Kim (1998) developed an evaluation framework consisting of 12 forms of electronic learning, mentoring, and assisting, which was then used to characterize online instructors interaction styles such as social (and cognitive) acknowledgement, questioning, direct instruction, and modeling. Meier et al. (2008) introduce a coding system that includes six dimensions covering aspects of communication, joint information processing, coordination, relationship management, and motivation. This coding scheme has been used successfully to both categorize group interactions as well as provide feedback to teachers and students (Kahrimanis et al., 2009). Johnson and Johnson (1996) propose a model that suggests that the major types of behavior that occur within a collaborative learning environment are giving and receiving help, exchanging resources and information, explaining and elaborating information; sharing existing knowledge with other; giving and receiving feedback; challenging others contributions; advocating increased effort and perseverance among peers; engaging in small group skills; monitoring each other s efforts and contributions. Curtis and Lawson (2001) then used this list to develop a coding system that was designed to characterize behaviors associated with positive social interdependence, as opposed to those behaviors linked to a more individualistic and competitive learning environment (Curtis & Lawson, 2001). The coding scheme consists of 15 behaviors grouped into five behavior categories. The five categories and their corresponding behaviors are listed in Table 1. The planning behavior category indicates that the message contains a statement that relates to organizing work, initiating activities, or group skills. The contributing code is assigned to messages that give help, provide feedback, exchange resources, share programming knowledge, challenge others or explain one s position. Other collaborative behaviors are also noted such as seeking input and reflection. Conversations about social matters that are unrelated to the group task at hand are placed in the social interaction category (Curtis & Lawson, 2001). Curtis and Lawson s coding scheme was utilized in the present study to explore the nature of collaborative behaviors that occur within global software development student projects. Since this study was interested in both categorizing and improving the communication among groups, it was felt that the Curtis and Lawson Table 1 Coding scheme and behavior categories (Curtis & Lawson, 2001). Behavior categories Planning Contributing Seeking input Reflection/monitoring Social interaction coding schema provided an accurate model of the types of behaviors that should occur within global software teams. The schema was used to determine which components of collaborative learning occur in the online interactions of students placed in global software teams. We applied this coding scheme to the logs of students chats and discussion forums. 3. Methodology 3.1. Overall design of the study Behaviors Group skills, GS Organizing work, OW Initiating activities, IA Help giving, HeG Feedback giving, FBG Exchanging resources and information, RI Sharing knowledge, SK Challenging others, Ch Explaining or elaborating, Ex Help seeking, HeS Feedback seeking, FBS Advocating effort, Ef Monitoring group effort, ME Reflecting on medium, RM Social interaction, SI Data for the study was obtained from three global software development student projects that occurred in fall and spring semesters in 2007 2008 and 2008 2009. The different global software teams were formed from students enrolled in courses at Atılım University (AU, Turkey), Panama Technology University (PTU, Panama) and University of North Texas (UNT, US). All three projects focused on topics related to software engineering problems. The projects were designed and developed by researchers at the three universities who were also teaching courses related to the different content areas in the software design process. The researchers determined the overall requirements of the programming assignments, as well as how the different projects would be integrated into existing curriculum. Thus, the actual programming assignments tended to vary according to the skill levels of the participants and the specific courses that were participating in the study for that semester. At the beginning of each project, students received training on how to use the communication software. Following the training, students were introduced to their team members (either through a teleconference or synchronous chat) and were provided information about the task as well as the management of their teams. Students enrolled in these courses received about 10 percent credit as part of their overall course grade for completing the project. To further motivate team participation, students were also awarded prizes for their participation and performance. For each assignment, student teams were asked to communicate with members through an open source learning management system, called OLAT. This system supports asynchronous communication tools such as forums, e-mails, file sharing, wiki discussion, etc. and synchronous communication such as chat. Since this system records the teams interactions, we are able to capture the communication behaviors for each team. A few teams also used external synchronous communication tools such as Skype, but these transcripts were saved and uploaded to OLAT folders.

494 F.C. Serçe et al. / Computers in Human Behavior 27 (2011) 490 503 3.2. Participants A total of 218 students participated in the three projects, all of whom were enrolled in either a computer science or information technology course at Atılım University (AU, Turkey), University of North Texas (UNT, US), or Panama Technology University (PTU, Panama). Table 2 provides a summary of the students who participated in each project. All of the 218 participants were between the ages of 19 and 25 years old. The average grade point average (GPA) for students in Panama and Turkey was around 2.0, while US students averaged 3.0. According to the students responses, 70% of the students stated that they had previously worked in a collaborative team, and only 1% of the students stated that they had never worked on a team project. The Turkey-based students were eight hours ahead of the US-based students and seven hours ahead of the Panama-based students. The US-based students were one hour ahead of the Panama-based students. 3.3. Global software development projects as collaborative tasks Research suggests that task type is one of the factors that may affect the type of communication behaviors that occur within a team. For example, some group tasks require collaboration only at the end when individually produced components must be merged together into a single product (Johnson & Johnson, 1996). Other tasks are straightforward, without generating any disagreements or misunderstandings (Dillenbourg & Schneider, 1995). Therefore, it is necessary that instructors are aware that different task features may affect both the amount and kind of collaboration that occurs among group members. The experimental collaborative tasks that were assigned to students in this study can be characterized as mid-size global software development projects. Global software development generally means that each remote team is assigned a different part of the software development process. For example, in a typical global software development project members located in one country may manage the project, while individuals in other countries take on the roles of testers, analyzers, coders, etc. Hence, global software development requires a great deal of communication among team members who are working on different parts of the project. Project managers, analysts, designers, testers, quality managers, configuration managers, etc., must all work collaboratively in order to produce a successful product. The specific roles assigned to each country for each project are listed in Table 3. Table 2 Participating Universities in each project. Project No. Turkey US Panama P AU UNT PTU Project 1 32 28 26 86 Project 2 38 29 12 79 Project 3 10 7 36 53 P 80 64 74 Table 3 Roles of students from AU, PTU and PTU in each project. Project 1 Project 2 Project 3 AU Java programmer Java programmer Java programmer PTU Algorithm developer Database designer Database designer UNT Web designer Database designer Leader Tester Leader Tester Three separate global software development projects were assigned to student groups who participated in this study. The specific software development tasks were determined, in part, by the content of the courses that were participating in the research for that semester. For example, if a research-faculty was teaching a programming course, then students enrolled in that particular course would be assigned the programmer role for the project. Similarly, if a research-faculty were teaching a design course, then the students would be the designers for the project. Participants were provided a description of the requirements for the programming task and a list of suggested assignments for teams in each country. A more detailed description of the three software development projects that were assigned during this study now follows. 3.3.1. Project 1 Student teams were assigned a code-intensive project that consisted of developing a system for creating and managing groups. Teams were asked to create a program that would allow a teacher to enter information about how teams should be configured and then execute the appropriate algorithm that would create teams according to specified input. This project s focus was only on the configuration portion of the team management system. Each country team was asked to deliver one of the components for the system and share that component with teammates in the other countries. The participating universities for this project were University of North Texas (UNT), Universidad Tecnológica de Panamá (PTU), and Atılım University (AU). The duration for the task was 6 weeks. This task was assigned to students in the spring semester of 2008. 3.3.2. Project 2 Student teams were asked to design, create and test a database software system that would manage a car rental system. This database development project included tasks for determining the functionality of the system, designing and implementing a test database, writing SQL queries, implementing java programs for testing, and producing a final report on the project. Each of these tasks described one of the deliverables. The dates and the tasks were pre-planned by the instructors. The participating universities for this project were UNT, PTU and AU. The duration for the task was 6 weeks. The task was assigned to students during the fall semester of 2009. 3.3.3. Project 3 Student teams were asked to design and develop a Book Store Management System that was located in Panama. Hence, students were to implement an interface that displayed Spanish text to the user. The teams were supposed to develop a stand-alone application in which the bookstore staff could add/update/remove the store s assets. The participating universities were UNT, PTU and AU. Unlike other projects, some of the researchers were also assigned roles as project leaders. The duration for the task was almost two months. The project was assigned to students in the spring semester of 2009. 3.4. Team composition There were a total of 34 virtual teams that participated in the three global software development projects. In the first project, 10 teams were created with approximately 3 students in each group from each of the three universities (for a total of 9 team members in each group). In the second project, 15 teams were created, with between 5 and 6 students in each group. Each team consisted of approximately 2 students from the US, 2 3 from Turkey, and 1 student from Panama. In the third-project, 9 teams were cre-

F.C. Serçe et al. / Computers in Human Behavior 27 (2011) 490 503 495 ated, with students from each university in each team. Table 4 summarizes the projects size and content, while Table 2 lists the roles for each country s teams. The students in Projects 1 and 2 were randomly assigned to their teams. This was also the case for Project 3, except for the role of team leader. The students were not allowed to change their team membership at any time during the project. The language for communication within the project teams was English, although the language for most of the co-located country teams in Panama and Turkey was their own native language. 3.5. Data collection techniques A combination of quantitative and qualitative research methods, including surveys and content analysis, was used in this study. At the beginning of each project, a survey was administered to team. This survey was designed to collect the demographic information about each student participant. Surveys were also used to collect information about students culture and attitudes about group work. In order to capture the interaction among the virtual teams and manage the different project deliverables, all teams were instructed to use the online communication tools to communicate. The open source platform learning management system called online learning and training (OLAT) was used to manage the group communications. This software supports asynchronous communications such as forums, e-mails, wikis, file sharing, etc., and synchronous communication such as chat. Data was obtained from the OLAT system directly, and from programs that were developed to augment OLAT s data collection capabilities. The data included information about a communication activity such as message posting, file upload, and wiki entry, along with the date, time, and author of each online activity. Whenever teams communicated through synchronous communication tools not supported by OLAT, there were asked to upload the logs to folder within their virtual space. A team s performance was evaluated by averaging the individual grades on each of the assignments. Projects were evaluated based on four criteria accuracy, efficiency, thoroughness, and style. A design or a program was considered accurate if it satisfied the user s functional requirements and contained no errors. A project s efficiency score was evaluated by examining the number of program modules. A program s thoroughness was scored on whether the design or program included all the necessary elements. Finally, good programming style was judged by examining the style (e.g., variable naming conventions, indentation, etc.) of the code. Researchers from each university graded their own student projects as well as those from the other participating countries. A mean grade for the project was then assigned to each student and team. to online discussion groups, online chats, and the teams performance grades on the projects. In order to explore the nature of team interaction in global software learning teams, each group s chat/forum discussion was coded to determine the overall number of communication behaviors devoted to planning, contributing, seeking input, reflection, and socializing. Since the research activities discussed in this paper were aimed at trying to characterize the collaborative behavior patterns within distributed software teams, we applied the coding scheme explained in Section 2.3. Two trained coders categorized messages into the 15 behaviors resulting in five behavior categories using the coding scheme. Each posting was extracted and coded into one of the communication behaviors. Codes were assigned to utterances in messages that indicated collaboration. Duplicate codes were assigned whenever an utterance indicated multiple collaborative behaviors. Unclassified messages that did not fit into any of the categories were also not counted. Percent agreement among the two coders for general content was 84.2%. 4.1. Communication sessions There was both asynchronous and/or synchronous communications in each project except in Project 2, which had only asynchronous communications. A total of 5298 messages were recorded during the three projects. Table 5 summarizes the distribution of both synchronous and asynchronous messages for each project. 4.2. Collaborative behaviors Codes, as defined by Curtis and Lawson (2001), were assigned to the 5298 utterances that occurred in the projects. Fig. 1 shows the Table 5 Projects and synchronous/asynchronous message profile. Project No. Chat messages Forum messages Project 1 2750 168 Project 2 136 Project 3 1531 713 P 4281 1017 4. Data analysis Data analysis generally involves the generation of data from the events or objects under investigation, the collection and maintenance of that data in records, and the transformation of the data into useful information. In this study, the data consisted of posts Fig. 1. Behavior categories and number of occurrences for all projects. Table 4 GSD projects and participants. Project No. Term Project title No. of teams No. of participants 1 Spring 2008 Group management 10 86 2 Fall 2009 Car rental system 15 79 3 Spring 2009 Bookstore for Spanish personnel 9 53

496 F.C. Serçe et al. / Computers in Human Behavior 27 (2011) 490 503 total number of collaborative behaviors that occurred for all groups and projects. Social interaction, feedback giving, and feedback seeking were the dominant behaviors that were observed. Figs. 2 4 show the collaborative behaviors that during each project. Figs. 2 4 show some obvious differences among the three projects. For example, Project 1 had the most number of communication behaviors in the social interaction category, whereas Project 3 had the most number in the feedback-giving category. Both Projects 1 and 3 had more group communications than Project 2, which seems to have had a much smaller number of communications and almost no social interactions. The chat and forum discussions within each project were analyzed separately in order to determine differences among behaviors in synchronous versus asynchronous modes of communication. Once each project s communication behaviors were separated, a cluster analysis was performed to determine patterns among the clusters. In this study, cluster analysis was performed to find the patterns of collaborative behaviors or characteristics Fig. 2. Behavior categories and number of occurrences in Project 1. among communications within projects. The groups or clusters that result from this classification process identified characteristics that maximally discriminate among the cases in different segments. The clustering variables were each group s number of interactions devoted to the five collaborative behaviors. Based on a review of clustering techniques, we chose a hybrid clustering method to identify the different groups. The hybrid clustering technique uses two methods namely K-means and Ward s hierarchical agglomerative clustering. The centers (or centroids) of each cluster are obtained first using Ward s method (Afifi, Clark, & May, 2004), a hierarchical cluster analysis technique that is said to be the most likely method to discover any underlying cluster structure. The resulting centroids are then used as the initial seed points for the non-hierarchical K-means cluster analysis. 4.3. Collaborative communication patterns in Project 1 The cluster analysis of the agglomeration schedule generated from Ward s method suggests a two-cluster solution for Project 1. Fig. 5 shows the total number of collaborative behaviors generated by each cluster for each communication category. Table 6 lists the specific groups assigned to each cluster. Each group is labeled by team number, communication source (i.e., chat or forum), and project number. For example, P1_1_Chat refers to the chat sessions generated by Team 1 in the first project, while P1_1_Forum refers to the forum discussion sessions of Team 1 in the first project. It s important to note that Table 6 shows that the clusters clearly differentiate between a group s chat versus forum postings. Fig. 5 shows that Clusters 1 and 2 have very different interaction patterns. For example, Cluster 1 (i.e., forum) has fewer overall collaborative behaviors than Cluster 2 (chat) as well as fewer collaborative behaviors in each communication category. To better understand the differences between the two clusters, we computed the proportion of communication behaviors devoted to each category (see Fig. 6). Results from a Chi-square test comparing the proportion of communication categories in the two clusters indicate significant differences between the two groups (v 2 = 139.97, df =4, p <.000). Post hoc tests indicate that the two clusters differed in the amount of planning (p <.0001), contributing (p <.005), seeking information (p <.0001), and social interaction (p <.0001) categories. There were no significant differences in the number of reflection behaviors between the two clusters. 4.4. Collaborative communication patterns in Project 2 Fig. 3. Behavior categories and number of occurrences in Project 2. The cluster analysis of the agglomeration schedule generated from Ward s method suggests two clusters for Project 2. Fig. 7 shows the number of collaborative behaviors in each cluster, while Table 7 lists the groups that were assigned to each cluster. It is Fig. 4. Behavior categories and number of occurrences in Project 3. Fig. 5. Number of communication behaviors in Project 1 by cluster.

F.C. Serçe et al. / Computers in Human Behavior 27 (2011) 490 503 497 Table 6 Teams assigned to each cluster in Project 1. Cluster No. Cluster 1 Cluster 2 Team P1_1_forum, P1_2_forum, P1_3_forum, P1_4_forum, P1_5_forum, P1_6_forum, P1_7_forum, P1_8_forum, P1_9_forum, P1_10_forum P1_3_chat, P1_1_chat, P1_2_chat, P1_4_chat, P1_6_chat, P1_7_chat, P1_10_chat, P1_5_chat, P1_8_chat, P1_9_chat important to remember that there were no synchronous communication sessions in Project 2; therefore, both clusters are comprised of groups whose only source of communication was the forum discussions. The total number of interactions among the team members is higher in Cluster 1 versus Cluster 2, although the total number of teams in Cluster 1 is much smaller than Cluster 2. Results from a Chi-square test indicate significant differences between the two clusters (v 2 = 10.33, df =4, p <.035). Follow-up tests show that there were significant differences between the proportion of planning (p <.02) and contributing (p <.003) behaviors between the two clusters. There were no differences between the two clusters in the proportion of Seeking Feedback, Reflection/Monitoring, or Social Interactions. In a previous study, we found that the differences between these two clusters were related to differences in performance. That is, high performing teams were assigned to Cluster 1 (Serce et al., 2009). 4.5. Collaborative communication patterns in Project 3 Fig. 6. Proportions of communication behaviors for each cluster for Project 1. The cluster analysis of the agglomeration schedule generated from Ward s method for Project 3 again suggests a two-cluster solution. Fig. 8 shows the clusters of collaborative behaviors observed in Project 3, and Table 8 lists the groups assigned to each cluster. Similar to the results found in Projects 1 and 2, the groups chat room communication behaviors tend to collect in Cluster 1, while the groups forum discussions (i.e., 89% of the communication messages) tend to gather in Cluster 2. Unlike previous projects, both Clusters 1 and 2 show similar patterns of collaborative behaviors. These similarities are shown more clearly in Fig. 9, which reports the proportion of communication behaviors devoted to each category for each cluster. Unlike the other projects, forum discussions appear to contain a more varied set of communication behaviors. Although the Chi-square test comparing the clusters showed that there were significant differences between the two clusters (v 2 = 11.14, df =4, p <.025), these differences occur only in the proportions for the contributing (p <.05) and social interaction (p <.001) categories. 4.6. Collaborative communication patterns in Projects 1 3 In order to better compare the behaviors in the three projects, clustering analysis was applied to all the data obtained from all three projects. The cluster analysis of the agglomeration schedule generated from Ward s method suggests three clusters. After using Fig. 7. Number of communication behaviors for Project 2 by cluster. Fig. 8. Number of communication behaviors for Project 3 by cluster. Table 7 Teams assigned to each cluster in Project 2. Cluster No. Cluster 1 Cluster 2 Team P2_1_forum, P2_11_forum, P2_15_forum P2_2_forum, P2_10_forum, P2_3_forum, P2_4_forum, P2_5_forum, P2_6_forum, P2_7_forum, P2_8_forum, P2_9_forum, P2_12_forum, P2_13_forum, P2_14_forum

498 F.C. Serçe et al. / Computers in Human Behavior 27 (2011) 490 503 Table 8 Teams assigned to each cluster in Project 3. Cluster No. Cluster 1 Cluster 2 Team P3_1_chat, P3_7_chat, P3_3_chat, P3_5_forum, P3_8_chat P3_3_forum, P3_1_forum, P3_2_forum, P3_4_forum, P3_6_forum, P3_7_forum, P3_8_forum, P3_6_chat, P3_9_forum Fig. 11. Proportion of communication behaviors for all projects by cluster. Fig. 9. Proportions of communication behaviors for each cluster for Project 3. Fig. 10. Number of communication behaviors for all projects by cluster. K-means clustering technique, the three clusters are generated as shown in Figs. 10 and 11. Table 9 lists the groups associated with each cluster. Again, the groups tend to be assigned to clusters based on mode of communication: chat behaviors assigned to Cluster 1; forum discussions assigned to Cluster 3; and a mix of behaviors assigned to Cluster 2. The data reported in Table 9 also seems to suggest that each project tends to generate its own communication patterns. For example, Cluster 1 appears to consist entirely of the communication behaviors generated in Project 1. Similarly, Cluster 2 is composed of (almost entirely) the communication generated in Project 3, whereas Cluster 3 is made up of the behaviors from all three projects. The total number of communication behaviors for each cluster is reported in Fig. 10. Cluster 1 appears to have the most number of communications, followed by Clusters 2 and 3. Fig. 11 shows the proportion of communication behaviors in each category for each cluster. A Chi-square test shows significant differences in the communication patterns among the three clusters (v 2 = 924.9633, df =8, p =.001). Table 10 shows the results of followup tests comparing the portion of behaviors in each communication category and cluster. Cluster 3 generated significantly more planning behaviors than either Cluster 1 or 2. All three clusters differ in their proportions of contributing behaviors. In contrast, the teams in all three clusters exhibited similar seeking input behaviors. Teams in Clusters 2 and 3 show similar social interaction behaviors, and both clusters differ significantly from teams in Cluster 3. Communication behaviors assigned to Cluster 1 were generated primarily from chats that occurred in a synchronous mode of communication. However, Cluster 2 includes behaviors obtained from both synchronous and asynchronous modes, while Cluster 3 consists of mostly asynchronous communication behaviors. 4.7. Language factor Since the participants language skills may have impacted both the type and amount of communication, we also looked at the communication patterns that occurred within each country. Research indicates that language and culture are highly related; thus, it was important to determine whether different types of communication behaviors can be associated with groups living in a particular country or region (Krutchen, 2004). As a result, the utterances of behaviors for each university were also applied to cluster analysis. Behaviors in each country s teams were separated out from the three-country teams and a cluster analysis was performed on the data. The Ward method suggests two clusters. After utilizing K-means clustering method, the resulting clusters are reported in Fig. 12. Cluster 1 includes 10 AU teams, 13 PTU teams and 3 UNT teams. Cluster 2 consists of 24 AU teams, 16 PTU teams and 40 UNT teams. The obvious difference between Clusters 1 and 2 is the number of utterances of collaborative behaviors in each category. The overall level of communication in Cluster 2 is much lower than Cluster 1. The unusually high number of UNT teams in Cluster 2 seems to imply that there are differences in communication behaviors among the three countries; however, a Chi-square comparing the proportions of communication behaviors between the two clusters was not statistically significant (as illustrated by Fig. 13). We then looked to see if there were any significant differences among the three-country teams within each cluster. We found significant differences among the teams in the three countries only within Cluster 1. The differences were primarily between Panama and the other two countries teams. Fig. 14 shows the proportion of communication behaviors for each country team in Cluster 1. There were differences between AU and PTY in the number of planning behaviors (v 2 = 12.49, df =2,p =.001), and in the number of reflection behaviors (v 2 = 20.83, df =2,p =.001). There were no differences among the country teams in the seeking behavior category. There were, however, differences between AU and PTY and

F.C. Serçe et al. / Computers in Human Behavior 27 (2011) 490 503 499 Table 9 Teams assigned to each cluster in Project 1. Cluster No. Cluster 1 Cluster 2 Cluster 3 Team P1_1_chat, P1_4_chat, P1_7_chat, P1_10_chat, P1_7_forum, P1_2_chat, P1_6_chat, P1_8_chat, P1_5_chat, P1_9_chat, P1_7_chat P1_3_chat, P2_11_forum, P3_1_forum, P3_2_forum, P3_3_forum, P3_4_forum, P3_5_forum, P3_6_forum, P3_7_forum, P3_8_forum, P3_9_forum, P3_1_chat, P3_3_chat, P3_8_chat P1_1_forum, P1_4_forum, P1_10_forum, P2_3_forum, P2_6_forum, P2_9_forum, P2_12_forum, P2_15_forum, P3_6_chat, P1_3_forum, P1_9_forum, P2_5_forum, P2_14_forum, P1_2_forum, P1_8_forum, P2_4_forum, P2_10_forum, P1_6_forum, P2_8_forum, P2_1_forum, P2_13_forum, P2_2_forum, P1_5_forum, P2_7_forum Table 10 Pairwise comparison of proportion of behaviors for each cluster. Communication behaviors Clusters 1 2 3 Planning [1,3] * [2, 3] * Contributing [1,2] ** [1,3] * [2,3] * Seeking input Reflection [1,2] ** Social interaction [1,2] * [1,3] The number in brackets indicates which cluster pairs are significantly different. * Significance at 99% confidence. ** Significance at 95% confidence. Fig. 14. Comparison of proportions of collaborative behaviors among country teams in Cluster 1. Fig. 12. Comparison of collaborative behaviors between two clusters. Fig. 15. Comparison of proportions of collaborative behaviors among country teams in Cluster 2. 5. Summary of the findings as a whole Fig. 13. Comparison of proportions of collaborative behaviors between Cluster 1 and 2. between PTY and UNT in the contributing (v 2 = 86.55, df =2, p =.001; v 2 = 60.31, df =2, p =.001) and social interaction (v 2 = 109.09, df =2,p =.001; v 2 = 46.49, df =2,p =.001) categories. Although Cluster 2 appears to show similar differences among the teams (see Fig. 15), none of the comparisons of the proportions among the communication behaviors for each country team was statistically significant. In the study, three separate global software collaborative development projects were undertaken with students enrolled in three different universities. The tasks assigned to global teams were chosen from the field of global software development. Each task concerned some elements of software design and development. Moreover, each task occurred within the context of a course assignment. Students were asked to use a particular online communication platform throughout the project. At the beginning of each task, the platform was introduced to the students. During the study, all communication among the distributed team members was logged. Qualitative content analysis methods were applied and K-means clustering with the Ward method was implemented in order to obtain detailed information about the patterns of collaborative behaviors observed within each project and among the projects as a whole. In this section, the results are summarized in terms of how different modes of communication