In the late 1980s, the World Health Organization promulgated a classification for the status of HIV transmission within countries: pattern I countries, characterized by spread among men who have sex with men, bisexual men, and injecting drug users; pattern II countries, whose spread was driven by heterosexual contact; and pattern III countries, whose low level endemic activity was related to their contact with countries of the first two patterns . Heterogeneity within countries, rapidly changing global forces, and undesirable national labeling, influenced World Health Organization to abandon this schema in the early 1990s in favor of a more complex classification called Global Areas of Affinity that encompassed epidemiologic, programmatic, and societal influences .
Both of these systems considered the relative impact of the major modes of HIV transmission (blood products, homosexual/bisexual, heterosexual, injecting drug use, maternal–infant transfer) but did not consider the potential impact of joint risk, that is, the simultaneity of multiple risk factors in an individual or in groups. Most classification schemes for HIV/AIDS risk, beginning with the Centers for Disease Control and Prevention's hierarchical classification, drew clear boundaries between risk groups. Among the first 1000 cases in the USA, for example, 727 were classified as ‘homosexual or bisexual men', although 81 of them were also ‘intravenous drug abusers’. Overlap was recognized, but downplayed, perhaps because of the considerable force of transmission generated independently by anal intercourse and needle-sharing.
The HIV/AIDS epidemic in the USA has evolved dramatically in recent years, with marked decreases in both incidence and mortality [4,5]. Attention has turned from rapid epidemic spread to persistent endemic transmission, primarily in urban centers. In this study, urban networks of persons are described whose risk-taking behavior and social network structure provide the substrate for ongoing endemic transmission of HIV, and whose entrenched transmission dynamics may characterize a significant portion of the epidemic in the USA for the foreseeable future. This analysis stems from the evolving hypothesis that individual behaviors do not sufficiently explain the propagation of infection, and network structure and dynamics play a critical, explanatory role by placing those individual behaviors in a context [6,7]. What is described here is a transmission system – the configuration of risks and interactions within a social network – that can be compared to other systems for their relationship to the transmission of infection.
The Urban Network Project developed from earlier work in Colorado Springs, Colorado, USA [7–12], an area of low HIV prevalence and low endemic transmission. A major concern of that study was reconciling the presence of groups at substantial risk for HIV (commercial sex workers, injecting drug users, and the sexual partners of people in these groups) with the infrequency of transmission. The Urban Network Project sought to examine similar issues in an area of higher prevalence and, presumably, of higher incidence for HIV. For this purpose, a similar conceptual approach was used for data collection and analysis, though the studies differed in the details of design and information ascertainment. A companion study, the Rural Network Project, has also been completed and will the subject of a separate report .
This longitudinal study enrolled respondents from 1995 through 1998, with follow-up through August 1999. The initial interview was followed by up to three subsequent interviews at 6 monthly intervals. The study design was motivated by a desire to capture the network configuration in which risky behavior took place, and to examine the dynamic changes that occur in such structure. After a period of ethnographic investigation, three geographically separated community sites were selected (designated sites A, B, and C) and two persons at each site were identified as ‘seeds’ or initial respondents to start a connected chain of persons and their contacts (Fig. 1). This type of sampling is necessitated by the hidden or elusive nature of the population at risk, a characteristic that precludes ‘top down’ or ‘descending’ sampling . Though a variety of names is given to the type of ethnographic sampling used  they all have a common theme: enumeration of a representative group based on a general knowledge of the community .
After interviewing an initial respondent, the second individual in the chain was selected from the first's contacts, the third from the second's, with continuation of this process until 10 persons and their contacts had been enrolled in the community chain. For one of the chains at each community site, the next respondent was chosen randomly from the contacts of the previous respondent (`random walk') . For the other, the contact was nominated by the respondent (`chain link') . For every chain constructed, all the contacts to the first, fourth, seventh, and tenth respondent were also interviewed. The net result, at the end of the first round of interviews, was the establishment of five community chains of 10 persons each that included their contacts, approximately 40% of whom were also interviewed. The sixth chain (site A, chain link) consisted of only seven persons, owing to a high level of recursion (persons naming those already named) in this group. Because of variability in the number of contacts, the number of respondents in the six groups varied from 10 at site A (chain link) to 36 at site B (chain link). By design, each of these community chains initially consisted of a single component of connected persons. It was anticipated that the single component would fragment with each reinterview, and so an attempt was made to interview not only the original respondents (not all of whom would be available) but also persons previously named and newly named by respondents nos. 1, 4, 7, and 10.
Virtually all interviewing took place in the target communities, at sites of the respondents’ choosing, or of convenience (a quiet part of the park; the interviewer's car) or, on occasion, at the county jail or other institutional site. Security issues were minimized by the frequent presence of project staff in the community, and the occasional service (referral, transport) that staff could provide for participants and others in the community. Ongoing ethnographic activity (not reported here in detail) also improved the community rapport needed to maintain contact with participants. Despite the urban facade, many community members narrowly defined their relevant social space. Although Atlanta is a large city with almost four million residents, most respondents in this study did not often leave their immediate neighborhoods. In one extreme example, a respondent had not left a half-mile radius from his home in more than 2 years. Despite these favorable conditions, follow-up was often made difficult by the purposeful or enforced ‘disappearance’ of participants in response to their life events.
The same instrument was used for the baseline and follow-up interviews, with a core set of questions divided into four main sections (survey instrument available on request). The questionnaire was designed to collect typical epidemiologic information as well as to maximize information on personal network structure. The questionnaire design was an integral part of the sampling strategy, by identifying persons to be interviewed at the next stage. The first section contained psychosocial and demographic variables such as place of residence, perceptions of the community, daily routine, and a brief medical history. Section two covered drug-related HIV risk, including an in-depth drug use history that collected information on the types of drugs used throughout the respondent's lifetime, as well as frequency of recent use. Section three addressed self-reported sexual behaviors, such as number of partners and frequency of oral, vaginal, and anal sex. The fourth section included a social network ascertainment, in which respondents provide the names, general locations, and descriptive background information of people with whom they associated. Contacts were defined as social (persons with whom the respondent shared resources such as food and shelter), sexual, drug-using, and needle-sharing. Friends and family members who did not fit any of these categories were occasionally mentioned, but were not systematically sought. A ‘matrix’ section, adopted from that used in the Colorado Springs study , requested information from respondents about all the relationships between all possible pairs of their contacts. A similar approach has been used in the General Social Survey, a large collection of egocentric information . A test for HIV was administered, after consent and appropriate pre-test counseling (followed by post-test counseling when the results were communicated to the respondent). ELISA and Western blot testing on blood was used initially, but the oral HIV test (Orasure, Epitope, Inc., Beaverton, Oregon, USA) was adopted when the technology became available, since phlebotomy was often difficult to perform. If the respondent was found to be HIV-positive, blood was collected to measure CD4 cell counts.
SAS  was used to perform standard epidemiologic analyses and extract data in a format usable by UCINET-V  and KRACKPLOT  for social network analysis. Incidence was calculated from the number of observed seroconversions divided by the total person-years of observation. For each of the first three semi-annual interviews in the six community chains, network properties were calculated for the total group (all multiplex connections) as well as the subgroups formed by sexual, social, drug-using, needle-sharing, and sexual plus needle-sharing connections (a total of 108 subgroups). Four main types of network properties were measured: (i) density, the number of actual connections divided by the total number of possible connections, provides an estimate of the interconnectedness of the network; (ii) centrality (both individual and network centralization), is measured using several different approaches and provides information on persons who may be critical in transmission and the extent to which a network is centered on such persons; (iii) weak components (number and the largest size component as a proportion of the total group), which convey information about the potential for spread though interconnectedness within a group; and (iv) measures of network microstructure (including cliques, k-cores, k-plexes) that provide a measure of the potential intensity of transmission through opportunities for an infectious agent to ‘cycle’ within small subgroups .
To assess the amount of interaction within a group, a variant of density that we called ‘recursion’ was calculated. Standard density is the number of actual connections (`e') divided by the number possible [N (N − 1) / 2], where N is the total number of persons (respondents and contacts) in the subgroup. Recursion is defined as the number of actual connections divided by the minimum possible for respondents only (`n'), given a particular study design. For this design, the minimum number of possible connections is n-1 (in a chain link of 10 persons, the minimum number of connections is nine; see Fig. 1). Thus, the minimum density is (n − 1) / [n (n − 1) / 2], or 2 / n, and recursion is actual density for respondents: e / [n (n − 1) / 2] divided by minimum density, or e / (n − 1). By comparing the actual density within a group to the minimum density possible given the study design, a simple measure of the degree of connectivity within the group is obtained. It is expected, for example, that a group at risk will have a high degree of interactivity, and will thus have recursion that is substantially greater than one.
In addition, concurrence (kappa) was calculated for each group, defined as the mean number of concurrent partnerships per partnership, according to the method of Kretzchmar and Morris, in which kappa is calculated as: κ = σ2 / μ − 1 where σ2 and μ are the variance and mean of the degree distribution in the subgraph [23–25].
Using a previously described method , the risk configuration faced by a person within a community chain was calculated as the number of known risky contacts (sexual and drug-using) divided by the number of multiplex relationships available (for each contact, this denominator is four: sexual, social, drug-using, or needle-sharing). Risk configuration represents the general level of risk in the network environment of an individual. To calculate risk behavior, the number of potential risky contacts (sexual + needle-sharing + (0.5) × drug-using + (0.1) × social) divided by the number of actual contacts that a respondent named was calculated. The constants (0.5 and 0.1) are empirical estimates that are used to correct for possible information loss. Risk behavior is a measure of the actual risks being taken by an individual.
To compare the frequency of network structures with those obtained in other studies, data were standardized by dividing the number of structures identified in a subgroup by the number of persons in the subgroup. Data from this study were compared to those reported previously (with some additional reanalysis) from Colorado Springs  by examining the ratio of the two frequencies, which we called the Adjusted Frequency Ratio. The differences in distribution of the actual adjusted frequencies for each of the structures measured were then tested using the Mann–Whitney test. For each microstructure measured in the subgraphs the 15 observations from Colorado Springs and the 15 observations from Atlanta were rank ordered and the smaller sum of ranks was compared to the Mann–Whitney T statistic .
A total of 292 persons were interviewed as part of the study. Of these, 228 were members of the six community chains, and the remainder were ‘isolates', persons who were thought to be ethnographically important, or who were believed to be connected to one of the chains, but had not been specifically named. Of the 228, 157 were part of the original recruitment into the six chains. Additional persons were added during subsequent interviews as new contacts were named by respondents. After the original 157 interviews, there were 133 interviews in the second wave (85%), 122 in the third (77%), and 110 in the fourth (70%). These percentages are inflated, however, by the new persons who were added at each wave. Among the 157 persons who were first interviewed, 119 (76%) were interviewed at least twice and 90 (57%) were interviewed a third time. Among the entire group of 228 participants, 72% were interviewed at least twice.
Within the six community chains, 62% of respondents were men, and on average persons of both sexes were about 40 years old (Table 1). Ninety-one percent were over the age of 30 years and 89% were African American (data not shown). Fifty-two percent of persons stated that they had not completed high school (48% of men and 54% of women), 54% were unemployed (49% of men and 61% of women), and 35% of men and 24% of women described themselves as homeless (Table 1). Perhaps more important, 80% of the entire group had one or more of these characteristics: 18.9% were both homeless and unemployed, and 17.5% were unemployed and had dropped out of high school. The prevalence of all of these characteristics varied markedly within the community chains; homelessness, for example, varied from 8% to 64%, and unemployment, from 33% to 88% (100% of the group with one person), in part perhaps because of the relatively small numbers of respondents in each group. Few persons self-identified as prostitutes or drug dealers. As noted in the next sections, considerable exchange of money for sex and drugs did take place, but is not captured by stereotypic classifications of roles.
Incidence and prevalence
In the entire group of 292 different persons, 234 (81%) were tested for HIV, and 31 (13.3%) were positive. Three seroconversions were observed during 164.8 person-years of observation, for an annualized incidence density of 1.8%. Based on the relationship that prevalence equals incidence times duration, the calculated mean duration of infection was 7.4 years, a result in keeping with the observed distribution of CD4 counts among infected persons. Overall, fewer than 5% of HIV-positive persons in this study had a CD4 cell count < 200 × 106/l. This proportion increased, however, from 1.7% at the first and second interviews to 9.0% at the third interview. This study was initiated at about the time that highly active antiretroviral therapy first became available, and only a few respondents in the study had access to it during the study period.
Patterns of drug use
The overall level of drug use in this population was high (Table 2). Alcohol and marijuana had been used by most persons, with introduction in the teen years. Crack cocaine was the predominant ‘hard’ drug, but the age at onset was considerably higher (28.7 years). Nearly one-fifth of the population had injected heroin within the last 6 months and 23% had injected cocaine. On average, persons injected themselves with drugs on nearly half the days of the preceding month. Opiates and amphetamines had been tried by a substantial portion of the population, but were not a major feature of drug activity.
Sexual and drug risks
Though men and women reported different levels of drug and sexual risk-taking, levels for both sexes were high (Table 3). Men reported over 25 sex contacts in the past 6 months, and nearly seven, on average, during the past 30 days. Comparable figures for women were substantially higher (81.3 and 14.6); women reported that most of these encounters were of a commercial nature, though, as noted, few classified themselves as prostitutes. In addition to the usual risk factors, such as women exchanging sex for drugs (71.2%), or the sharing of needles with multiple persons by both men and women, several unanticipated factors were observed. Though only 6.9% of men stated that they had a male sexual orientation, 9.4% reported that they had sex with men, and 28.3% reported practicing insertive anal intercourse. In addition, 38.6% of men reported that they had received money for sex, and 35.4% reported that they had received drugs for sex – two practices reported more frequently among women. Nearly 19% of women stated that they had a female sexual orientation, and one-third reported that they had had sex with women.
Respondents identified both single and multiple relationships with their partners (Table 4). Uniquely sexual or needle-sharing relationships were uncommon in all the community chains (< 10%), but non-needle-sharing drug use relationships alone occurred commonly (20–40%). Simultaneous sexual and needle-sharing relationships between the respondent and a contact were not common (also < 10%), but combined sexual and non-needle-sharing drug use relationships occurred more frequently. The patterns of multiplex relationship vary considerably among the six community chains. For example, 32.8% of relationships at site A (chain link) involved both needle-sharing and drug use, compared to only 3.4% of relationships at site C (random walk). The relative absence of combined sexual and social relationships in any of the groups suggests that pairing in domestic relationships is infrequent, a finding confirmed by ethnographic observations.
Adjusted frequency ratio
Though not true in every case, the frequency of network structures (e.g. cliques, k-plexes, etc.) was in general substantially higher in Atlanta than in the Colorado Springs study . When comparing the subgroups with any relationships, or each of the four possible relationships individually (Table 5), five of the ratios were undefined because the value in both Atlanta or Colorado Springs was zero; in two the ratio was 1.0; in eight of the categories the ratio was less than one (Colorado Springs > Atlanta) and for the remaining 45 categories, the ratio was greater than one (Atlanta > Colorado Springs). A Mann–Whitney test  on the rank order of the actual frequencies for the two sites for each of the four microstructures using all five possible relationships demonstrated that three of the four were significant at the P < 0.01 level and the fourth narrowly missed significance at the P < 0.05 level.
General network properties
Six measures were used to summarize the interactions of network structure and risk-taking (Table 6). The largest component within the subgroups of sexual and needle-sharing connections included between 17% and 94% of group members. The distribution was bimodal: half of the groups were over 50% and the other half were under 35%. Risk configuration (the amount of potential risk faced by a member of the network) was uniformly < 20%, with an average over the 18 observations of 11%. Risk configuration uniformly declined over the period of observations, though most of these declines were small. Risk behavior (a measure of the actual risks incurred by participants based on the nature of their relationships) was higher than risk configuration in all groups, indicating a substantial level of risk taking as a result of the multiplex relationships within networks. Risk behavior also declined in all groups over the period of observation but with only one exception (site C, chain link), these declines were small. Despite the substantial level of same-sex sexual activity by men within this overall population, same-sex partners (as a percentage of all sex partners) did not figure prominently in any of the chains.
Kappa, a measure of the number of concurrent partnerships per partnership, is substantial in the first interview in each community chain (varying from 3.27 to 9.62), but declines thereafter at all sites. This finding suggests that the simultaneity of sexual or needle-sharing contact diminished over the period of observation. Recursion, which compares actual density among those interviewed to minimal density based on the study design, was measured only for the first interview, during which a single connected component was constructed. Recursion was substantial in all but one group, indicating a density up to 2.5 times higher than the minimum expected. An overall look of all the measures included in Table 6 provide little that is unique about the two community chains [site A (chain link) and site B (chain link)] in which seroconversions occurred. An overview of the data also suggests that there is little to distinguish between the random walk and chain link approaches.
Though hierarchical classification schemes and risk-specific intervention programs have been in place for some time, there is considerable current interest in the more complex nature of risk [27–33] but little systematic documentation of such complexity. The picture derived from this quantitative overview of the urban networks of persons at risk for HIV is striking for both multiplicity within groups and heterogeneity between groups. Multiplicity is evident in the frequency with which persons engage in a variety of risk behaviors (drug-using; sex risk, including frequent contact with persons of the same and opposite sex; and surrogates, such as exchanging drugs for money, that have been shown to be associated with transmission); in the frequency of multiple relationships between partners (multiplexity, in network parlance); and in the frequency with which multiple risks exist within the same network. The striking level of same-sex partnerships, in the absence of self-identification as gay or lesbian, is general evidence for the polymorphous nature of these interactions. In this setting of inner city, minority persons at risk for HIV, the application of hierarchical or univariate classification of risk would impede an understanding of transmission dynamics.
Though the subjective judgment is debatable, we have described the level of prevalence (13.3%) and incidence (1.8% per year) as moderate. In these communities, we have also described the observed joint risk as moderate and stable (or declining) in most groups examined, and marked by considerable heterogeneity. The terminology stems from a comparison with Colorado Springs, previously classified as a low risk area , based on a prevalence in the groups at risk of approximately 3%, and the virtual absence of transmission over the course of the study. Only one seroconversion in a contact (none in respondents) was observed during the study interval  In this comparison, the usual questionnaire-based methods for assessing risk produced similar results (see Table 3 and ). Using a network-based assessment, risk configuration in the two Colorado Springs cohorts varied from 17 to 20% (higher than the risk configuration mean of 11% reported for this study), but risk behavior in Colorado Springs varied only from 10 to 20%, compared to an average of 50% in Atlanta. The ratio of risk behavior in Atlanta compared to Colorado Springs thus ranged from 2.5 to 5.0. Similarly, the frequency of microstructures in Atlanta was substantially greater than that in Colorado Springs, with an adjusted frequency ratio averaging 8.0 for non-zero values. Network-based quantitation of risk and structure cannot be related to disease occurrence directly in this study, but the prevalence ratio of 4.4 (13.3% / 3.0%) is consistent with the network findings. Though comparable data are not available for a high prevalence and incidence area, the fact that transmission probabilities are generally low  leads to the hypothesis that a considerable level of interactive joint activity is required to maintain higher levels of endemicity. One caveat in the interpretation of these findings is that somewhat different sampling schemes were used in Colorado Springs and Atlanta. Both, however, attempted to maximize the likelihood of finding interconnectivity, the latter through a purposeful design, the former through recruitment of persons (`cross-links') named by several others. Finally, we do not have longer term data that could describe the epidemiologic picture at the time (7–8 years ago) when many in this study might have contracted their infection. The underlying assumption, based on 4 years of follow up (data not shown) is that personnel change, but the underlying structure is not dramatically altered.
Two recent large-scale interventions obtained differing results when they tested the use of community-based sexually transmitted disease treatment as a mechanism for diminishing the transmission of HIV. The trial in Mwanza, Tanzania  observed baselines prevalences of 3.8% and 4.4% in intervention and comparison communities and a cumulative probabilities of HIV-seroconversion of 1.2% and 1.9%. In Rakai, Uganda, with a baseline HIV prevalence of 15.9%, a seroconversion rate of 1.5% per year was observed in both intervention and comparison groups . The differences are the subject of intense, ongoing debate, but of interest here is that the incidence in both areas was similar to that observed in the Atlanta community groups, and the prevalence in this study approximated that in Rakai. Our study suggests the possibility that network factors may have played a role in the differential impact of sexually transmitted disease treatment on HIV transmission. In addition, it readjusts our perception of the intensity of HIV transmission in some communities in the USA.
This study highlights not only the multiplicity of risk interactions, but also the stereotypic classifications that are used to describe persons at risk. Traditional designations such as ‘heterosexual', ‘homosexual', ‘prostitute', ‘drug user', ‘drug dealer', and similar appelations were of little use in these communities because of the frequency of multiple roles, and the changing patterns of risk taking. Though relatively few men self-identified as ‘homosexual', a substantial number reported having sex with men, including insertive anal intercourse. Similarly, only a few women in the study defined themselves as working prostitutes, though the majority had exchanged sex for money or drugs. The heterogeneity of risk within groups, and the changing patterns of risk over the period of observations suggests that individuals’ activity also varies in the short term.
One result of this complex pattern was that prediction of individual seroconversion from network structure and behavioral patterns is difficult. The three persons who were observed to seroconvert, and their networks, could not be readily distinguished from others. Thus, the pattern described by this study, together with earlier work, offers a general blueprint for the level and intensity of endemic transmission, but is difficult to use as a predictive model. It may be that some combination of mathematical modeling in conjunction with the continued collection of this type of data would be required to provide a firmer quantitative basis for linking incidence and prevalence with network structure and risk behavior. Such assessment might serve as the basis for targeting intervention services to areas and groups in greatest need.
The authors thank R. Anso, C. Smith and the staff of the Urban Networks Project for their commitment to this work; and R. Needle and E. Lambert for their support and encouragement. In addition, the comments of anonymous reviewers were most helpful in shaping the manuscript.
1. Chin J, Mann JM. The global patterns and prevalence of AIDS and HIV infection.
AIDS 1988, 2 (suppl 1): S247 –S252.
2. Mann J, Tarantola DJM, Netter TW. AIDS in the World.
Cambridge MA: Harvard University Press; 1992.
3. Jaffe HW, Bregman DJ, Selik RM. Acquired immune deficiency syndrome in the United States: the first 1000 cases
. J Infect Dis 1983, 148: 339 –345.
4. Centers for Disease Control and Prevention. HIV/AIDS Surveillance Report. 1999, 11: 32 –37.
5. Palella FJ, Delaney KM, Moorman AC. et al. Declining morbidity and mortality among patients with advanced human immunodeficiency virus infection
. N Engl J Med 1998, 338: 853 –860.
6. Aral SO. Sexual network patterns as determinants of STD rates: paradigm shift in the behavioral epidemiology of STDs made visible
. Sex Trans Dis 1999, 26: 262 –264.
7. Rothenberg RB, Potterat JJ, Woodhouse DE, Muth SQ, Darrow WW, Klovdahl AS. Social network dynamics and HIV transmission
. AIDS 1998, 12: 1529 –1536.
8. Klovdahl AS, Potterat JJ, Woodhouse DE, Muth JB, Muth SQ, Darrow WW. Social networks and infectious disease: The Colorado Springs Study.
Soc Sci Med 1994, 38: 79 –88.
9. Potterat JJ, Woodhouse DE, Rothenberg RR. et al. AIDS in Colorado Springs: is there an epidemic?
AIDS 1993, 7: 1517 –1521.
10. Rothenberg RB, Woodhouse DE, Potterat JJ, Muth SQ, Darrow WW, Klovdahl AS. Social networks in disease transmission: The Colorado Springs Study.
In Social Networks, Drug Abuse, and HIV Transmission. NIDA Research Monograph151
. Edited by Needle RH, Coyle SL, Genser SG, Trotter RT. Rockville: Department of Health and Human Services, Public Health Service, National Institutes of Health; 1995: 3 –19.
11. Rothenberg RB, Potterat JJ, Woodhouse DE, Darrow WW, Muth SQ, Klovdahl AS. Choosing a centrality measure: Epidemiologic correlates in the Colorado Springs study of social networks
. Social Networks 1995, 17: 273 –297.
12. Woodhouse DE, Rothenberg RR, Potterat JJ. et al. Mapping a social network of heterosexuals at high risk for human immunodeficiency virus infection
. AIDS 1994, 8: 1331 –1336.
13. Trotter R, Rothenberg RB, Coyle S. Drug abuse and HIV prevention research: expanding paradigms and network contributions to risk reduction
. Connections 1996, 18: 29 –45.
14. Spreen M. Rare populations, hidden populations, and link-tracing designs: what and why?
Bull Methodologie Sociologique 1992, 36: 34 –58.
15. Rothenberg RB. Commentary: sampling in social networks
. Connections 1995, 18: 105 –111.
16. Watters JK, Biernacki P. Targeted sampling: options for the study of hidden populations.
Soc Prob 1989, 36: 416 –430.
17. Klovdahl AS. Social networks in an urban area: first Canberra study
. Austral New Zeal J Sociol 1977, 13: 169 –175.
18. Burt RS. A note on the general social survey's ersatz network density item.
Social Networks 1987, 9: 75 –85.
19. SAS Institute Inc. SAS Language and Procedures: Usage, Version 6, First Edition.
Cary, North Carolina: SAS Institute Inc.; 1989.
20. Borgatti SP, Everett M, Freeman L. Ucinet V for Windows: Software for Social Network Analysis.
Natick, Massachusetts: Analytic Technologies; 1999
21. Krackhardt D, Blythe J, McGrath C. KRACKPLOT 3.0
. Pittsburgh, Pennsylvania: Analytic Technologies; 1995.
22. Wasserman S, Faust K. Social Network Analysis.
Melbourne: Cambridge University Press; 1994.
23. Kretzchmar M, Morris M. Measures of concurrency in networks and the spread of infectious disease.
Mathemat Biosci 1996, 133: 165 –195.
24. Morris M, Kretzschmar M. Concurrent partnerships and transmission dynamics in networks.
Social Networks 1995, 17: 299 –318.
25. Morris M, Kretzchmar M. Concurrent partnerships and the spread of HIV.
AIDS 1997, 11: 641 –648.
26. Snedecor GW, Cochran WG. Statistical Methods.
Ames, Iowa: The Iowa State University Press; 1972.
27. Chu A, Torbert A, Brown L, et al.Overlapping AIDS epidemic among injecting drug use (IDU) and heterosexual exposure (HT) in New York City (NYC) AIDS cases
. Fifth Conference on Retroviruses and Opportunistic Infections
. Chicago, February 1998 [abstract 129].
28. Klevens RM, Fleming PL, Mays M. Patterns of reporting multiple risks for HIV infection in the United States
. Fourth Conference on Retroviruses and Opportunistic Infections.
Washington DC, January 1997 [abstract 495].
29. MacQueen KM, Ciesielski C. Race/ethnicity differences among homosexual/bisexual men with AIDS who have multiple modes for HIV transmission
. IX International Conference on AIDS
. Berlin 1993 [abstract PO-D06-3591].
30. Harwell TS, Trino R, Rudy B, Yorkman S, Gollub EL. Sexual activity, substance use, and HIV/STD knowledge among detained male adolescents with multiple versus first admissions
. Sex Trans Dis 1999, 5: 265 –271.
31. Perlman DC, Henman AR, Kochems L, Paone D, Salomon N, Des Jarlais D. Doing a shotgun: a drug use practice and its relationship to sexual behaviors and infection risk
. Soc Sci Med 1999, 10: 1441 –1448.
32. Elifson KW, Boles J, Darrow WW, Sterk CE. HIV seroprevalence and risk factors among clients of female and male prostitutes
. J Acquir Immune Def Syndr 1999, 2: 195 –200.
33. Logan TK, Leukefeld C, Farabee D. Sexual and drug use behaviors among women crack users: implications for prevention
. AIDS Educ Prevent 1998, 4: 327 –340.
34. Potterat JJ, Rothenberg RB, Muth SQ. Network structural dynamics and infectious disease propagation
. Int J STD AIDS 1999, 10: 182 –185.
35. Royce RA, Sena A, Cates W, Cohen MS. Sexual transmission of HIV
. N Engl J Med 1997, 336: 1072 –1078.
36. Grosskurth H, Mosha F, Todd J. et al. Impact of improved treatment of sexually transmitted idseases on HIV infection in rural Tanzania: randomised controlled trial
. Lancet 1995, 346: 530 –536.
37. Wawer MJ, Sewankambo NK, Serwadda D. et al. Control of sexually transmitted diseases for AIDS prevention in Uganda: a randomised community trial
. Lancet 1999, 353: 525 –535.