Although there are a wide range of conceptualizations of social capital 17 , we take a network perspective, which holds that the extent to which an individual can realize the benefits of social capital is a function of that individual's position in a given social network 10 18 . This drives our focus on: a) the resources that can be accessed by network members (either directly or through contacts), and b) the structure of relationships or linkages in a network of interest 10 . In a professional network, key benefits of increased social capital among colleagues include increased exchange of information and resources 17 19 . For example, sharing of appropriate and timely information allows individuals to make strategic adjustments to reach their goals 10 20 . Additionally, participants can access novel information by developing relationships with individuals who are dissimilar in terms of experience and professional contacts 21 . By learning in the context of social relationships, network members can come together to identify pressing problems, make sense of complex changes in the environment, and develop innovative solutions 22 23 . Provision of tangible support or material resources from one network member to another also improves network members' performance 24 . By tapping into relationships, network members can gain access to contacts' resources, and perhaps more importantly, to the resources held by the organization(s) represented by those contacts 25 . The challenge is to balance efficiency (knowing others who have contacts and resources that are very different than one's own) and effectiveness (development of a strong set of key contacts) 18 .

Social network analysis provides the necessary tools for our analysis as the methodology allows for the assessment of structures in social relationships, as well as the resources exchanged through those relationships 26 . Additionally, given that successful capacity-building relies on changes at the individual, organizational, and system levels 27 , the ability to assess relationships and resource flow at multiple levels allows for a holistic assessment. For example, a network in which all members are connected prompts members to develop trust and a sense of obligation towards each other and encourages the generation of social capital 28 . At the same time, at the individual level, connections to other network members are expected to provide new access to resources for program participants. If a capacity-building program results in network structures that support resource exchange, network-based social capital can have an impact on the ultimate goal of management training programs: the improvement of trainee performance.

Despite the number of programs focused on building healthcare worker capacity 2 7 8 9 and the understanding that increased collaboration and partnerships are important outcomes of capacity-building efforts 29 , we are not aware of previous studies examining how such programs may affect the structure and functioning of resulting social networks. Examining this potential impact is important to our understanding of the full impact of capacity-building programs in health. Using survey data from hospital executives participating in an executive-education program in Ethiopia 30 , we conducted a social network analysis to examine the growth of the network and the social capital generated by the network (in the form of resource exchange) during the first year of the program. Social network development and social capital generation were not explicit goals of the training program, but we were able to take advantage of this natural experiment to test exploratory hypotheses. We expected to find growth and resource exchange within networks as well as a positive association between network connections and resource exchange. We tested these assumptions among a network of program participants and among a network of participants plus educators and mentors participating in the program.

The capacity-building program under study was a two-year executive-education Master of Hospital and Healthcare Administration (MHA) program in Ethiopia developed by the Federal Ministry of Health (FMOH), the Clinton HIV/AIDS Initiative (CHAI), Jimma University, and the Yale School of Public Health 9 31 . The program was implemented at the request of the FMOH, with the goal of developing skilled executives to improve hospital management in Ethiopia, a low-resource, high-demand setting. This program was part of a larger quality improvement effort targeting the Ethiopian healthcare system, which began decentralization in 1994. The course was offered by Jimma University in Jimma, Ethiopia and was the first graduate-level program for hospital management in the country. The course was administered and taught jointly by faculty from Jimma and Yale Universities, with local coordination provided by a Program Director and Program Assistant. As an executive-education program, the course was offered over two years, with three-week long sessions in residence three times per year, as well as regular progress reports and evaluations when trainees were working at their hospitals.

Executives of public hospitals were eligible to apply. The course focused on improving trainees' skills in a range of management-related areas, such as human resources, hospital operations, financial management, strategic planning, and leadership. Trainees also had the opportunity to develop professional connections with each other as well as with leaders and mentors in Ethiopia and the United States.

We conducted a cross-sectional study at the end of the first year of the MHA program to describe the social networks that developed during the year. Data were collected with a self-administered survey of two groups of respondents: trainees and supporters. Trainees were the first Chief Executive Officers (CEOs) of public hospitals in Ethiopia. Supporters comprised educators and mentors formally linked with the MHA program through either Yale or Jimma University or through CHAI. We contacted all 25 trainees enrolled in the MHA program and 38 supporters affiliated with the program to complete the survey. All research procedures were approved by the Human Investigation Committee at the Yale School of Public Health and the Institutional Review Board at Jimma University.

The self-administered survey was distributed in December 2008 and January 2009 and required approximately 20 minutes to complete. Paper copies of the survey were distributed to all trainees in residence during the December course session and electronic copies were distributed to all other respondents. Surveys were administered in English, which was the language of instruction and a requirement for participants in the MHA program.

For this study, we focused on two networks: 1) the Trainee Network, which was comprised solely of trainees, and 2) the Trainee-Supporter Network, which included trainees and supporters (educators and mentors). Respondents were presented with a roster that listed all trainees and supporters. The survey asked all respondents to identify trainees and supporters with whom they interacted for professional purposes. Respondents also noted whether or not they were acquainted with each network member before the MHA program started. From these responses, we derived our measures of interest for each network.

We measured a series of network characteristics which have been shown in other settings to promote exchange of information and flow through networks 26 . These measures were based on data about connections (or reported relationships) between network members. Some measures focus on presence or absence of a connection, whereas others include information about the 'direction' of the connection. For the latter, the measure can capture whether Member X reported a connection to Member Y, Y reported a connection to X, or both reported a connection to each other.

To describe the network as a whole, the first measure of interest was network density, or the proportion of possible relationships between members that were realized, which described the extent to which network members are connected, regardless of the direction of connections 26 . A more dense, or more highly connected, network may be useful for sharing information and resources and cooperation, whereas a more sparsely connected network may provide greater access to diverse contacts and novel resources 10 18 . A density level of around 15-20% is expected to support knowledge-sharing in a network of about 100 members 32 . We also identified isolates, individuals who reported no connections to other network members. Isolates are of interest as their lack of connections prevents them from contributing to or benefiting from network membership. Last, we identified components, or subgroups of members that are not connected to each other and therefore cannot share information and resources between subgroups 26 .

Shifting our focus to individual network members, we calculated degree, the number of connections between a given network member and all other network members, regardless of the direction of ties 33 . The bulk of our analyses focused on out-degree, or connections from a given network member to other network members. Thus, if Member X reported three connections with other network members, that member's out-degree value would be three, regardless of how many network members reported connections to Member X. Compared with degree, this measure narrows the focus to connections that may be perceived as functionally useful to respondents 34 ; here, these connections involve the set of individuals from whom respondents may seek and gain skills. In the Trainee Network, 'trainee out-degree' was the number of connections a trainee reported regarding other trainees, grouped into tertiles. In the Trainee-Supporter Network, 'trainee-supporter out-degree' was the number of connections to supporters reported by each trainee, grouped into tertiles. Last, geographic homophily referred to whether or not pairs of network members worked in the same region.

To assess potential by-products of social network development, we measured informational and functional exchanges, which are complementary manifestations of social capital that can help trainees achieve work-related goals 10 24 . Informational exchange refers to access to necessary knowledge, the ability to transmit it to the correct person, and acquisition of information with sufficient time to react 18 . Trainees were asked whether or not they received guidance in non-classroom settings from: a) other trainees, and b) supporters on a series of subjects. These topics included: problem-solving, human resources, finance management and budgeting, basic public health, biostatistics/research methods, hospital operations, strategic management, health policy development and analysis, health ethics and public health law, leadership, and management information or tools. The list of topics was defined in the curriculum as critical to the program and most topics, but not all, were covered in the MHA course at the time of the survey. We created a summary score of the total number of exchanges reported and dichotomized responses at the 50^th percentile for each network, resulting in categories of 'low exchange' and 'high exchange' for each network. Based on the distribution of data, 'low exchange represents zero reported informational exchange in the Trainee Network.

Functional exchange described the provision of tangible support from one network member to another 24 . Such exchange often involves collaboration between institutions or individuals that benefit one party to a greater degree, e.g., one individual training another on the use of a new tool. Examples of tangible support can include sharing of useful tools, policies, and materials or serving as a reference for colleagues 25 35 . Trainees were asked whether or not they received a series of tangible resources from: a) other trainees, and b) supporters. These resources included: materials and goods (such as surplus supplies), connections/introductions, and hands-on instruction, such as through site visits. We created a summary score of the total number of exchanges reported and dichotomized responses at the 50^th percentile for each network, resulting in categories of 'low exchange' and 'high exchange' for each network.

We conducted a sociometric network analysis for both the 25-member trainee network and for the larger 63-member trainee-supporter network, which included educators and mentors (n = 38) in addition to trainees (n = 25). Sociometric analyses assess the connections between all members of each network of interest, supporting evaluation of network growth and resource exchange 36 37 . Thus, an individual who was invited to participate, but did not fill out a survey, could have been noted as a contact by another respondent and would still appear in the dataset. Although the Trainee Network is wholly contained within the Trainee-Supporter Network, we analyzed them separately to be able to isolate resource exchange among complementary sets of ties that are important for trainees.

Network analysis requires dedicated software to assess relational data, and we used UCINET-6 38 . As network data are not independent and do not meet the assumptions of classical statistical techniques, we utilized procedures developed for network data available in the UCINET software package 38 39 . Thus, the significance tests were based on random permutations of matrices as is appropriate for relational data. Here, the significance levels were determined based on distributions created from 10 000 random permutations. The analytic procedures also supported comparison of matrices of data. Descriptive measures were calculated using standard UCINET procedures developed for network data. We utilized UCINET Contingency Table Analysis to assess the association of out-degree with two types of resource exchange. We tested the relationship between geographic homophily and connection patterns using UCINET QAP Relational Cross-Tabulation.

Among trainees, 23 of 25 individuals completed the survey (92% response rate). Table 1 describes the characteristics of trainees' hospitals. The trainee hospitals had on average 204 beds with a range of 40-800 beds, and the average number of employees per hospital was 399 employees, with a range of 82-2500 employees. The majority of hospitals (72%) were classified as regional; one-third were rural.

For the larger network, 41 of 63 individuals completed the survey (65% response rate), with a 47% response rate among supporters. Network-level growth was assessed using a pair of network graphs (Figure 2) and a series of complementary measures (Table 4). The density increased from 3% to 13% of all potential ties realized over the first year of the program. We analyzed density increases among subgroups and found increased ties from trainees to supporters (3% to 20%), from supporters to trainees (0% to 12%) and from supporters to supporters (5% to 9%). The number of isolates decreased from 8 to 2 in this network, and there was only one component at year 1, ignoring isolates. Again, increased variation in out-degree and in-degree values for the full network from the beginning of the program to year 1 suggests that the network became more centralized. Assessing the overall network, the individuals with the most connections in this network were mainly faculty and staff that played a central role in program administration and teaching.

When we narrowed our focus to relationships between trainees and supporters, we found that at year 1, 94% of trainees reported informational exchange with supporters and 55% reported functional exchange with supporters. The average trainee-supporter out-degree at year 1 was 8.1 connections. In this network, the average number of outgoing connections with supporters was 2.3 for the lowest trainee-supporter out-degree tertile, 5.3 for the middle tertile, and 14.9 for the highest tertile. Trainee-supporter out-degree did not vary significantly between regions.

As seen in Table 5, trainee-supporter out-degree was positively associated with informational exchange, χ²(1, N = 23) = 74.93, p < 0.05. Those in the highest tertile of trainee-supporter out-degree also had the highest reports of informational exchange. We found a similar pattern for trainee-supporter out-degree and functional exchange, χ²(1, N = 23) = 81.31, p < 0.01.