Local Network Studies

Three teams and two contributed papers used local network designs. They showed that a great deal of rich data can be obtained from local network designs and much can be learned about the spread of HIV.

The National Health and Social Life Survey (NHSLS) and Chicago Health and Social Life Survey (CHSLS)

The first large-scale local network study to be conducted on sexual networks was the NHSLS. The NHSLS was conducted in 1992 and a follow-on study, the CHSLS was conducted from 1995 to 1997. Both studies collected extensive information on local sexual networks as well as the practices that occurred within these relations. They employed face-to-face interviews on samples of women and men between 18 and 59 years of age. The NHSLS used a national household probability sample, and the CHSLS drew samples at two geographic levels: the city level, which includes an inner suburban ring of Chicago, and four targeted-neighbourhood areas within the City of Chicago.

The response rate was 78.6% for the NHSLS and ranged from 60% to 78% for the Chicago samples. In addition to interviewer-administered questions, both surveys included self-administered sections to improve response rates on potentially sensitive topics such as same-gender sexual experiences and drug use. CHSLS features recent, local sexual and social friendship networks, while the NHSLS had more extensive information about lifetime sexual networks.

For each respondent the NHSLS included information on sexual behaviour, basic demographic information on up to 28 lifetime sexual partners, and detailed demographic and sexual information on up to two last-year partners. The NHSLS also gathered self-reported data on health-related characteristics, including nine common sexually transmitted diseases (STDs). The CHSLS asked many of the same basic questions as the NHSLS, but also included new sections on social networks (up to six confidants of respondents), geographic location, and relationship characteristics for the respondent's two most recent sexual partners.

A number of important findings have been published from these studies regarding the transmission of HIV/AIDS. The data have been used to examine determinants and implications of partner concurrency and the use of condoms. Concurrency has been shown to increase the speed of disease transmission in populations. The findings show extensive cultural variation in concurrent partnership patterns among groups. The data have also been used to develop a network-based epidemiological model that explains differential rates of STDs by race/ethnicity. While the overall rates of the traditional bacterial STDs have been declining in the uS, the decline has been much more rapid for whites than for members of other race and ethnic groups. These data suggest that the differentials persist in part due to the structure of sexual networks: a high level of assortative mixing by race, coupled to lower "core group" segregation among non-whites. The Chicago data also show that having a strong social network is associated with lower rates of infection, even among persons with the same number of sex partners.

The Thai and ugandan Sexual Network Studies

Martina Morris presented on two local sexual network studies: one from Thailand (from the Behavioural Research for AIDS Prevention study) and the other from uganda (the ugandan Sexual Network and Behaviour Study). The design of both surveys was strongly influenced by the NHSLS study, as Dr. Morris had worked on the NHSLS during its design phase. The Thai survey was conducted in 1992-3 and is the first study to collect local sexual network data outside the uS. It employed quota samples from three populations: low income men (15-49 years old), long-haul truckers (15 years old and up), and women working in inexpensive brothels (12-53 years old). The ugandan survey was conducted in 1994, and used a stratified household-based random sample of men and women 15-45 years old from 90 communities in the Rakai district of uganda. The communities were stratified into rural, intermediate and trading centre villages.

Both surveys used the same name generator -- the person you had sex with most recently (and then the two partners previous to this) -- collecting information on up to three most recent sexual partners. The ugandan survey also collected information on the first sexual partner. Information included attributes of the partners, attributes of the relationship, and the partner-specific behavioural repertoire. Dr. Morris said she learned to ask for specific dates when asking when the first and last sex occurred, and if the relationship was ongoing, to ask specifically, "Do you expect to have sex again." Both studies also conducted focus groups and in-depth interviews.

The data from these studies have been used to examine a wide range of questions at the individual and population level. The Thai study has been used to estimate the impact of men as a "bridge population", linking women who work in the sex industry to women who do not. The ugandan data have been used to examine the geographic network created by sexual partnerships. The findings here suggest that the links between the higher prevalence urban and trading centres to the rural areas are much more likely to take the form of a sequence of links across more proximate villages, rather than a single direct link. These data have also been used to drive simulations to understand the role of concurrency in the transmission dynamics of HIV. Concurrency creates larger connected components, and increase the speed and pervasiveness of HIV spread. Morris noted that concurrency measures should be regarded as indicators for population level risk not individual risk. For the individual, having concurrent partners is simply a form of having multiple partners, there is no additional risk incurred by having the partners simultaneously. It is the partners who are put at risk if the individual has them concurrently, as they are now exposed indirectly to an additional person. The classic example is a monogamous woman whose male partner has other partners. Thus, we should not expect to see concurrency to be a significant risk factor in a typical logistic regression on serostatus. But we would expect to see that communities with high levels of concurrency would have higher prevalence, all other things equal.

Finally, the detailed pair-specific behaviour has revealed both within- and between- person differences in risk taking. Travellers in uganda were found to be twice as likely to use condoms as non-travellers, suggesting travellers might play a role in disease prevention, as well as spread. Sex workers and male clients in Thailand were found to be much less likely to use condoms with their "regular" commercial partners, evidence that relational dynamics are often more important than knowledge and information in determining condom use. The comparability in the two surveys makes it possible to do substantial comparative work in each of these areas, highlighting the variations in the social organisation of sexuality, and their impact on the population dynamics of HIV and other STIs.

The WHO/uNAIDS 4 Cities Study

The third set of studies to use a local network design was The WHO/uNAIDS 4 Cities Study. These are a set of population based studies in four African towns, conducted between July 1997 and March 1998. The study sites were selected on the basis of differential HIV prevalence among pregnant women and trends in prevalence over time. As high HIV prevalence towns Kisumu (Kenya) and Ndola (Zambia) were selected. The low HIV prevalence towns are Cotonou (Benin) and Yaoundé (Cameroon). One set of respondents came from a household probability sample of men and women aged 15 to 40 years. The target sample size in each site was 1,000 men and 1,000 women, and response rates were good. A team consisting of interviewers and nurses or doctors visited each household. Study participants were interviewed on their socio-demographic characteristics and sexual behaviour, using a standardised questionnaire, and were asked to provide blood and urine specimens. The other set of respondents came from a representative sample of about 300 CSWs. These were interviewed and also asked to provide specimens.

Examination of the trends of HIV prevalence over time suggests that the current differences in prevalence are due to differences in the rate of spread of the virus rather than differences in time since the start of the epidemic. Therefore, the compelling question behind the study was what explains the variation in the prevalence levels of HIV? Two broad explanations were examined: sexual behaviour patterns (probability of exposure to an infected partner) and factors that determine the probability of HIV transmission per sex act (kind of sexual behaviour, co-factors such as STDs, lack of male circumcision and high viral load of the infecting partner).

Rates of concurrency at the aggregate level did not correlate to the levels of HIV prevalence in the 4 cities, nor was there an association between concurrent partnerships and HIV infection at the individual level. The lack of an individual level correlation was not surprising, as concurrency is more likely to put the partners at risk than the focal respondent. But the lack of association at the community level was puzzling. In these sites, as elsewhere in sub-Saharan Africa, women are much more likely than men to be HIV positive at ages 15-24. Three possible explanations were considered: Data artefacts, women are more likely to have infected partners, and women are more susceptible. Once again, it appears that biological explanations predominate, and that women are more susceptible. The preliminary conclusions of the studies were that current reported sexual behaviour alone cannot explain the variations in the epidemic in the four cities. Younger age at first intercourse of women, young age at first marriage and large age differences between spouses made a difference. But, biological differences, or factors influencing the transmission of HIV, seem to be driving the different epidemics.

Contributed Papers

Two local network analyses were presented in the contributed paper sessions. Simon Gregson also examined the higher HIV prevalence among young women in rural Zimbabwe, again attempting to test hypotheses of biological versus behavioural determinants. In contrast to the 4 Cities Study, however, Gregson concludes that the partnerships that young women have provide far greater exposure to HIV, a behavioural argument. Among young women, coitus is more frequent, condom use is sporadic and their partners, being older and more likely to be working, are more likely to be infected. Women’s partners probably also are more likely to have other partners. Thus, the different locations of young men and women within local sexual networks and differences in the nature of their sexual experiences within relationships result in very different levels of exposure to HIV. This structure is supported by the social, cultural and economic structure, which are not easy to change. Soori Nnko and Basia Zaba presented findings from a local sexual network study in Kisesa, Tanzania. A survey of all adults 15-44 years was carried out in 1994/95 and repeated in 1996/97. A sexual network module was added at the end of the interview during the second survey (adapted from the uNAIDS module for sexual networks, similar to the one used in the 4 Cities Study). Fifty eight percent of sexually active men and 22% sexually active women reported at least one non-marital, non-cohabiting partnership. Among those who reported at least one partnership, the average number of partnerships was 2.1 for men and 1.2 for women.

Each of these local network studies have shown that asking questions about sexual networks did not lead to low response rates or interview breakoffs.

Partial Network Studies

Three research teams presented work using partial network designs. Each team used some form of contract tracing in their study design, and reported that they had good response rates and were able to collect good quality data.

The Colorado Springs Study

John Potterat, Donald Woodhouse and Steve Muth presented findings from the Colorado Springs Study. This study focused on the networks of heterosexuals at high risk of HIV in a city with several large military bases, substantial commercial sex and drug use, and significant STD prevalence. The study employed design an open-cohort longitudinal design and data were collected from 1987-1991. The primary sampling frame consisted of prostitute women, injecting drug users, and their respective sexual and drug partners. Respondents were recruited from health department STD, HIV, and substance abuse clinics; vice squad referrals and street outreach. Respondents were asked to give full names and locating information of social and family contacts, sex partners and drug partners for the 6-month period preceding an interview. This enabled the investigators to employ some snowball sampling (which define the study as a partial network design), tracing and attempting to enrol partners who had been named by two or more participants. All respondents were also asked to provide a blood sample for HIV, syphilis and hepatitis testing. Of 1079 eligible persons, 595 (55%) were enrolled. Over 5 years, they completed 990 interviews, and named nearly 7,000 persons as contacts or associates (in a community with a population of 400,000). About two-thirds of the contacts could be uniquely identified, and half of the remaining cases are known not to be duplicates. Enrolment was not strictly random, but the authors believed it to be representative.

This study provides one of the most comprehensive samples of a relatively bounded network in the field of HIV research. It has supplied some of the first images and analyses we have of the road surface that HIV travels on. The network is highly connected in Colorado Springs, with over 300 of the 595 respondents joined in a single connected component. While this suggests that HIV should have spread rapidly through the network, that did not happen. The study suggests that the reason is because the few HIV-positive persons found during the study period (less than 20) were either not part of this large component, or were on its margins.

Trust and rapport between the interviewers and respondents was key in the high amount of co-operation found in this seemingly invasive research. No participant elected to withdraw from the study due to the blood testing and only four people walked out of interviews.

Atlanta and Flagstaff Network Studies

These studies were based largely on the lessons learned (and questions generated by) the Colorado Springs Study. Hoping to look at the interrelationship of drug-related and sex-related risk for HIV in two very different communities, the authors collected partial network data on social/sexual/drug-using networks. They developed an elaborate snowball sampling scheme to obtain comparable network information in the two settings. In each community they established two community "chains" of persons. Each chain was started by interviewing an initial person (the "seed"). The seed was asked to identify all of their drug, sex or social contacts. One of these contacts was chosen to be the second link in "the chain" being formed (randomly chosen in half the cases – called a random walk technique, and by seed nomination in the other half – called a chain link technique). The third link in the chain was chosen from the contacts named by this second link, and so forth to obtain a total of ten links. In addition, all the contacts to persons in link positions 1, 4, 7 and 10 in the chain were interviewed. In the course of the study, the investigators also interviewed a number of persons who were not explicitly named as a contact by any of the respondents, but who, based on ethnographic data, were known to be central in the life of the community. The final sample size was 292 persons, of whom 226 were members of community chains, and these respondents were followed over a 3-year period. At least one follow-up interview was obtained with 76% of the 226 respondents in the community chains. Like the Colorado Studies, they found informants very forthcoming with information.

The Seattle Studies

This group of studies was designed to examine the sexual mixing patterns, sexual networks and specific types of sexual partnerships in Seattle populations with and without sexually transmitted diseases. The samples include attendees at STD clinics; attendees at non-STD health facilities; partners of STD infected individuals reached through contact tracing; respondents recruited at socio-geographically specified areas of the city; as well as representative samples of the general population based on census tract of residence.

These partial network designs provided very rich data. The data have been used to examine mixing patterns by race, age and geographic areas, and to analyse the impact of mixing on the risk of infection. The authors also examined whether the impact of mixing differs by type of infection (e.g., bacterial vs. viral, or genital warts vs. genital herpes), and reverse the question and compare network composition across groups defined by infection status. As partner information is collected both from the original respondent and from the enrolled partners, the authors can compare the two reports. In addition to empirical analyses, the authors have used their data in a number of simulation studies in collaboration with Geoff Garnett.

Complete Network Designs

Complete network is perhaps a bit of a misnomer in the context of HIV transmission, as one should probably define the global population as the complete network at risk of HIV. There are two studies, however, that have collected saturation samples in certain bounded communities and examined behaviours that have relevance to HIV and STD transmission, and their teams were invited to present at this conference. The studies help to clarify the limitations of the other network designs, by showing what can be done when the whole network is obtained. They also provide a good reality check, as the costs and logistics of conducting such surveys are daunting.

The National Longitudinal Study of Adolescent Health

The National Longitudinal Study of Adolescent Health (Add Health) Study is a major milestone in survey design at many levels. The study is designed to assess the health status of adolescents in the united States and explore the effects of the multiple contexts (both social and physical) in which they live. The primary sampling frame was a national list of high schools. From this, a sample of 80 high schools (and in some cases, their feeder schools) was selected. The survey consisted of two parts: an "in-school" questionnaire, completed by all of the students in the sampled schools (more than 90,000 adolescents), and an "in-home" questionnaire, completed by a sample of students in each school (about 12,000 adolescents). For 16 "saturation" schools, all enrolled students were selected for both the in-school and in-home interviews (about 6,000 adolescents). This yields a nationally representative sample of adolescents.

The in-school questionnaire asked students to nominate up to five male and five female friends and to indicate which of several activities they had done with each of these friends during the past week. Because the nominations can be matched to student rosters, this yields nearly complete social network data for most schools. Friendship networks can be determined and a respondent’s peer group, as well as his or her position within it, can be described in detail. The in-home interviews went a step further, eliciting up to three of the students' romantic partners using Computer Assisted Survey Instruments, and Audio assisted CASI for sensitive questions. These complete network data are unparalleled in scope. They provide a unique look at the component structure of a complete network, and how it evolves over time. Results from one of the saturated schools suggest the romantic network structure, which is characterised by a large circular ring, could be generated by a simple norm that prohibits partner swapping (in network terms, a norm against 4-cycles).

Confidentiality was a major concern in this survey, and an elaborate security system was used. The team separated identities and responses and sent the identities to Canada so they could not be subpoenaed. The usefulness of this strategy was never tested. Deductive disclosure was also an issue, and access to the data, though PUBLIC, requires several levels of security clearance. Another the major barrier was resistance at every step, from parents’ groups, to teachers regarding classroom time and time of staff in school.

The Nang Rong Study

The Nang Rong Projects are a loosely integrated collection of studies designed to monitor and understand the sweeping demographic, social, and environmental changes that have taken place over the past twenty years in Nang Rong, Thailand. An interest in social networks in relation to fertility and contraceptive behaviour, migration processes, land use and social processes informs much of the work. These also represent some of the few, if not the only studies that combine both complete social network data and spatial data.

Social network data in the Nang Rong Projects come mainly from surveys administered in 1994 and 1995. There were three such surveys: a household survey (current and former household members in 51 villages); a migrant follow-up (out-migrants from 22 villages); and a community profile (310 villages). Another round of surveys is scheduled for 2000 and 2001, which will expand the scope and temporal depth of the social network data and enable an examination of network change over time.

The Nang Rong surveys provide information about multiple social relations, at multiple levels of observation. The village data include information on ties between villages based on sharing temples, schools, water sources, bus routes, and access to major highways as well as those arising directly from labour exchanges and equipment rental. The household data include information on ties to other households based on sibling relationships, help with the last rice harvest; the renting, hiring, and sharing of agricultural equipment, and prior residence as well as those arising indirectly from the use of local rice mills and membership in village organisations. The migrant data include information on contact with other migrants from the origin villages, visits and exchanges of money and goods with origin households, sibling ties and some ego-based questions about friends and acquaintances in the place of destination.

A distinctive feature of the Nang Rong data is the availability of data on complete population-based social networks. Many important measures of network structure and position, such as subgroups and their membership, indirect connections between network members via paths containing intermediaries, and some kinds of network positions can only be calculated from complete networks. Information about social ties was collected for all households in the study villages, and for all villages in Nang Rong District. Household and village identifiers are explicitly recorded. For households, the availability of complete data provides replicated social networks for 51 villages. A great strength of the data is the rich information on characteristics of households (e.g., size and composition, contraceptive choices of married women, agricultural activities) and villages (e.g., resources, social institutions, health care) that can be used to help interpret network patterns at both household and village level and to understand variation in network patterns across villages.

Other Contributed Papers

Statistical techniques to analyse network data are in their infancy, but many people are working on improving network measures, as was evidenced by the work presented. Many participants gave suggestions about what they would like network models to be able to do in the future.

Models and Methods

Geoff Garnett spoke clearly on an issue raised several times during the conference: the two functions of network research methods and statistics. The first function is to describe the actual network, and the second is to understand the implications of structure on infection. He noted that for the transmission of HIV, there are many factors involved in a network, such as patterns of contacts (partner change and choice), state of the individual (infectivity), and the strength of contact (transmission possibility). He believes that we need statistics that can describe the complexity that exists within networks. Models can and should include factors that affect transmission, such as condom use, circumcision, pill use, etc. Dr. Garnett noted that there is lots of work to be done, but that adding sexual networks and other factors to models will make them better reflect reality. However, increasing the detail of the data increases the cost in terms of the study and complexity. Therefore, one needs to weigh the benefits in increased understanding with the costs.

In the models that many participants would like to see, time will be a factor included. James Moody presented a paper that underscored the importance of relationship timing for STD diffusion. He explained that sexual networks have a peculiar temporal feature: while each dyadic relationship is symmetric, the timing of the indirect relations implied by partners of partners are not symmetric. Indirect asymmetry follows from the fact that infections can only flow forward through time-ordered paths. Thus, one’s past partners are not at risk to one’s current infection while one’s future partners are. This temporal ordering is a key element needed to understand disease diffusion in the population, as small changes in time ordering of relations have a huge effect on infection possibility. Depending on the timing of relations, the same contact network may result in dramatically different sets of people exposed to the STD. Dr. Moody noted that because concurrency equalises the timing of relations, it allows infections to flow both ways rather than one way. This helps to explain why concurrency increases the rate of spread.

Dr. Moody showed how the timing of sexual relations determines potential exposure and provides a method for deriving upper and lower bounds for indirect contact within networks of a known structure. He then applied the method to data on romantic and sexual networks in the united States, showing how relationship timing both limits the ultimate exposure within the network and creates pools of people who are mutually connected indirectly, thus forming pockets where an STD could circulate. This work implies that understanding STD diffusion requires collecting data on the timing as well as the pattern of sexual relations. He noted that p* models can take into account the direction and value of a tie. Therefore they could take into account the infectivity and direction of infectivity across ties. Thus, we can use these models to predict outcome for an individual (although it is hard to do) and can simulate a structure and run a simulation to tell how disease will spread. Identifying temporal patterns raises new methodological questions.

Most social network analysis packages ignore time, Dr. Moody suggested, with the exception of PAJEK. John Potterat of the Colorado studies also noted the capacity restrictions of network software as a barrier when the Colorado studies began. He noted that still the lack of bug-free software has been a problem for them.

The fact that difference aspect of networks are relevant for different question was also underscored several times in the conference. Dr. Garnett believes that the most important factors in the spread of HIV are the number of partners and concurrency. However, as Dr. Borgatti pointed out, concurrency is more important in the number of people a person infects, and the number of sex partners is important for both acquiring and transmitting HIV. James Moody made the same point when presenting findings of the Add Health Project. The implications of the ties of a complete network depend on the disease you are tracking, the order of ties, the concurrency, timing, etc. One needs to consider the biology of diseases as well as network structure.

Laura Koehly presented work on a set of models that provide a crosswalk between local network data and complete network data. The methods are based on random graph models for contact networks: loglinear models, logit models and P*. Log-linear models were introduced by Dr. Morris (1991) as a method for analysing data on assortative mixing by attributes (age, race, etc) that can be obtained from local network data. These models assume dyadic independence, and thus ignore micro-level relational structures. P* models, on the other hand, provide a more general framework for modelling partnership structure. This approach allows for more sophisticated modelling of the social interactions within attribute groups, while also accounting for attribute composition and mixing. The log-linear models can be shown as special cases of the more general random graph models represented by p*. The common exponential form makes it possible to use this modelling framework to investigate the role of network microstructures and aggregate mixing biases on disease transmission dynamics.

Dr. Koehly explained that p* was originally designed for complete network data, but that a link with log-linear models for local networks had been found. Work to adapt p* for partial networks analysis still needs to be done. Future research will also involve the development of longitudinal p* models that will allow us to investigate network change over time.

Linking Networks and Spatial Processes

Networks are structured by the likelihood of a link between two partners, and this likelihood is both a function of physical space, and the demographic attributes that define social distance. It is not surprising, therefore, that the statistical models for networks resemble the statistical models for spatial processes. One of the most promising areas for future research lies in the integration of social and physical space into a common modelling framework.. Two studies at this conference presented work that incorporated network findings into Geographic Information Systems (GIS). Jenna Mahay, using CHSLS data, described sexual partnership choice and its spatial mapping in Chicago. She showed that the sexual marketplace is a spatial place where people look for and find partners. It is not open to everyone, but is differentiated by social positions, such as class, language, culture or race. It is also differentiated spatially. Jenna suggested that the use of spatial and social mixing maps can be used to look at the potential geographic spread of disease.

The data from the Nang Rong Projects have also been incorporated into a GIS. Geographic information about village locations makes it possible to properly orient the graphs of the village networks and study variability in their spatial arrangement. Insights into patterns of network ties can be gained by mapping village networks in reference to land cover data derived from remote images. The spatial analytic capabilities of the GIS also make it possible to assess the impact of the administratively defined district boundary and to evaluate whether rivers and perennial streams create barriers to network ties between villages.

Another application of networks with implications for the spread of HIV is in looking at the causes and consequences of migration, and how HIV affects or is affected by migration. The Nang Rong Projects, as well and the local network study in uganda both considered migration issues in their research.

Networks of Information and Support

In studying the spread of HIV through networks, we were reminded that social networks could be a way to spread information. They can also exert social control and offer social support to reduce the spread of HIV. The social network information in the CHSLS is likely to be particularly useful for building intervention and prevention programmes since information about STDs is transmitted through both social and sexual networks.

Dr. Goyal also presented a paper on the role of discussion networks in diffusing information and behavioural change through social networks. He presented findings of an operations research study undertaken to assess the perception, attitude and behaviour of men and women towards sexual health, to understand their information needs on this subject, and to meet these needs through dialogue between men and women within the same social network. Against the backdrop of a low literacy level and limited exposure to mass-media, interpersonal communication, or dialogue, within and between the sexes proved to be more effective among members of same social network, because social pressures for attitudinal and behavioural modifications are stronger in such settings.

The project was implemented in one rural and one urban locality of Jaipur District, India. The study showed that both men and women had a very poor perception of their sexual health and sexual health needs. After initial hesitation, both men and women could frankly discuss sexual health and related issues in dialogue sessions with people of the same sex and eventually of both sexes. The community response to these initiatives was very positive, and there was a large and effective diffusion of ‘dialogue’ in the community at large.

The study shows that dialogue is an effective mode to promote sharing of perceptions and experiences among men and women. It is also effective in creating awareness and validating scientific knowledge. It is particularly helpful in understanding one’s own sexuality and appreciating the sexuality of the opposite sex. It also contributes to modifying opinions and attitudes of men and women towards safe sex. More importantly, dialogue has helped create a positive environment for adoption of safe sex practices and strengthen women’s ability to negotiate less violent and safer sex within marital unions.

A final application touched on in the conference was in the opening of the conference, when Dr. USA Duangsa, a resident of Chiang Mai who works for uNAIDS in Bangkok spoke. She encouraged participants to expand concern from the transmission of HIV to the social networks of caregivers and organisations providing care for those infected for the years of life they live with the virus. This concern was reiterated by Basia Zaba at the closing ceremony.

Data Quality

Dr. Morris said that one of the strengths of network studies is that they are more like a conversation than an interrogation about sexual practices, which makes respondents more comfortable with disclosure. They move in the direction of ethnographic research, where you can see how relationships evolve, build a person’s life history and collect important contextual data for the observed behaviour. She reported that the ugandan study had a 90% response rate with only four breakoffs in the whole study. She noted that it was remarkable what respondents will tell you just by asking about partners. Tony Pramualratana, an anthropologist on the Thai Network study, added that it was an art to develop the ability to conduct a conversational questionnaire, to establish rapport, to obtain responses, and to finish in 90 minutes. Done well, network surveys do not necessarily take longer than traditional ones.

Many presenters noted the rich quality data they collected. Pamina Gorbach discussed valuable qualitative data she is using to examine the motives that lead adolescents and adults into specific types of partnerships and networks. She used structured, in-depth interviews with 270 respondents. The volume of data is overwhelming and expensive to work with, but very rich. She is finding that the definitions of concepts such main partner might not be as specific as we might think. Sevgi Aral also pointed out there are many types of concurrent partners. The data being collected is clearly heightening our understandings.

The researchers involved in the Colorado Springs Studies have used interviewers who have been working in the community for many years. They believe this helped with the quality data they collected. The Atlanta/Flagstaff studies did not have that luxury, but did use outreach staff who were former drug users to help connect the study to the community. They concluded that if long-term field presence is not possible, the acquisition and training of persons who can conduct interviews with populations at high risk for HIV is the single most crucial element for success.

Alex Weinreb presented a paper that examines the effect of interviewers knowing their respondents in network studies that are focused on social relationships and personal behaviour. Three findings emerged. First, insider-interviewers were much more successful in getting data on network partners. Second, a substantial amount of response variance on critical dependent variables was associated with the difference in interviewer-stimulus (i.e. insiders vs. strangers, men vs. women). And third, the joint effect of these two factors in analysis leads to the estimation of radically different coefficients on certain network characteristics when the contribution of different types of interviewers to the data is emphasised. This implies that results are contingent on the type of interviewer that is used. This is certainly worth bearing in mind given the norm of using stranger-interviewers in survey research.

Dr. Weinreb argued that social network research is more prone to interviewer effects because the response effects are worse for network data. For example, there is more noise around an answer about a partner’s characteristic than one’s own. He found insiders collected more network partners and information. Insider men get more kinds of family planning information than females of any kind. He believes this is not due to a personal failing of an interviewer, but to the relationship between interviewer and respondent. It has to do with social relations, which network researchers are interested in anyway. He does not know who is getting correct information, just that it is different. This points to a need to study this more and to diversify interviewers. Respondents continued this debate during the final session, wondering about what characteristics make the best interviewers.

While many presenters noted rich data and high response rates, there were several who raised questions about the validity of the responses. Simon Gregson said that the research team he works with has the feeling people are not completely telling them the truth. Researchers in the Four Cities studies found the same. This may be more a function of the subject matter – sexual behaviour – than the survey design, but it cautions against assuming that network studies solve the disclosure issue.

The alter adjacency matrices may be particularly prone to reporting problems. Alter adjacency matrices are the data collected by asking respondents in a local network survey to tell you about the relationships among the partners they have reported. The Atlanta / Flagstaff studies conducted a supplemental study that compared respondent reports of the alter adjacency matrix with what their partners said about their own interactions with the same people. The preliminary results suggest that only about 12% of the relationships identified by respondents among their partners are verified by the partners themselves. At the same time, the resulting structural characteristics defined by the alter adjacency matrix is similar to that which emerges from the interviews with the partners.

Soori Nnko’s qualitative data from Kisesa, Tanzania shows that you can get rich sexual networking data. Still there was a lot of concern about data quality. Single women tend to underreport sexual partnerships, and women exaggerate the length of current partnerships, especially those with high status men. Even though women report much lower frequencies of multiple partnerships, preliminary analysis indicates that women's partnership patterns are potentially a more sensitive indicator of sexual networking in the population than the corresponding male data. This suggests that the women's’ data may be more reliable.

However, it was pointed out that data quality issues are not unique to network research. The advantage of network data is often that such problems, hidden in ordinary surveys, can be revealed by using the cross-validation techniques that network data collection make possible.

Hélène Voeten presented research conducted in the rural province of Nyanza in Kenya, where multiple methods were used to investigate the same population on the same or overlapping issues to gain insights into sexual networks and to validate self-reports from surveys. The focus of the research was on sexual behaviour patterns and networks of young adults aged 15-29, commercial sex workers and their clients. The following multiple methods were used: 1) household based study of young adults: questionnaires followed by in-depth interviews with a sub-set of the same respondents; 2) CSWs: focus group discussions followed by either a questionnaire or an in-depth interview, which were again followed by a 2-weeks sexual diary keeping; and 3) CSW clients: informal discussions mainly in bars. This study overlapped with the sites of the Four City Study, and the questionnaire was adapted from that study. In-depth interviews were conducted with a sub-sample of the same respondents.

Overall the consistency between data collected with questionnaires and in-depth interviews is fairly large and thus reliability of the collected data is rather high. Women proved to give more consistent answers than men. However, since there is no gold standard for measuring sexual behaviour, it is difficult to interpret results in terms of validity. Analyses of response patterns from both methods do not suggest much difference in recall and social desirability bias between the two methods. However, the fact that the majority of responses of males as well as females shift toward more risky behaviour during the in-depth interview suggests that recall bias for the in-depth interviews was smaller. Therefore, the data collected with in-depth interviews seem more valid. How much difference does this make in modelling and network analysis has not yet been determined.

Ethical Issues

Ethical issues came up several times during the conference. The primary ethical issue is the potential for the confidentiality of the sensitive information to be compromised. For the Colorado studies, Dr. Potterat explained that data management and information that was linked to names were major ethnical issues. They dealt with this better over the years of their research, using more sophisticated computer technology and creating sheets with identifying information that could be sent out of the country, to avoid potential subpoenas, much as the Add Health team did. The authors were committed to ensuring their research did no harm to respondents.

A second ethical issue was whether to separate interventions from data collection in the Colorado Springs Studies. They decided not to separate the two. For example, after each blood sample was tested, participants were informed of the results, and were counselled and given information about how to protect themselves. A similar procedure was used in the Atlanta/Flagstaff studies, offering pre- and post-counselling for blood tests and walking respondents through services. In Colorado, the interventions might have had an effect. At the start of the study, less than 5% of respondents used condoms, while presently 40% report condom use. This is enough of a change to do damage to bridges and disturb the HIV epidemic.

A further ethical issue was whether to offer payment for interviews. The Atlanta / Flagstaff interviewers grappled with giving money to people if they knew the respondents would use the money to buy drugs. They also struggled with situations where respondents who would name fictitious partners in exchange for payment. This debate continued during the training fellows session at the conclusion of the conference, where participants weighed giving something back to people who offer so much, but not wanting to offer gifts or money that would unduly influence respondents’ choice to participate. In uganda, the approach taken by the research team was to support the community health clinics, as a way to thank the entire community for support of the research. The benefits were thus made available to both persons who did and did not participate.

Kristen Velyvis