Article Text
Abstract
Objectives: A population-based sexual network study was used to identify sexual network structures associated with sexually transmitted infection (STI) risk, and to evaluate the degree to which the use of network-level data furthers the understanding of STI risk.
Methods: Participants (n = 655) were from the baseline and 12-month follow-up waves of a 2001–2 population-based longitudinal study of sexual networks among urban African–American adolescents. Sexual network position was characterised as the interaction between degree (number of partners) and two-reach centrality (number of partners’ partners), resulting in the following five positions: confirmed dyad, unconfirmed dyad, periphery of non-dyadic component, centre of star-like component and interior of non-star component. STI risk was measured as laboratory-confirmed infection with gonorrhoea and/or chlamydia.
Results: Results of logistic regression models with generalised estimating equations showed that being in the centre of a sexual network component increased the odds of infection at least sixfold compared with being in a confirmed dyad. Individuals on the periphery of non-dyadic components were nearly five times more likely to be infected than individuals in confirmed dyads, despite having only one partner. Measuring network position using only individual-based information led to twofold underestimates of the associations between STI risk and network position.
Conclusions: These results demonstrate the importance of measuring sexual network structure using network data to fully capture the probability of exposure to an infected partner.
Statistics from Altmetric.com
Risk of acquiring a sexually transmitted infection (STI) such as HIV is commonly thought to be determined by three factors: the rate of exposure to new partners, the probability that a partner is infected and the efficiency with which infection is transmitted from the infected to the uninfected individual.1 2 Despite the widespread acceptance of this framework, the probability that a partner is infected is rarely measured in epidemiological studies of STIs. We hypothesise that this is largely because of the difficulty in measuring that parameter, since individuals often do not know whether their partners are infected.3 4 5 However, for STIs, as well as for infectious diseases that are transmitted through person-to-person contact, collection of social network data offers an alternative method for estimating this important parameter.
Given that STIs are transmitted through sexual linkages, individuals’ and partners’ positions in their sexual networks can serve as proxies for the likelihood that a person will be exposed to an STI. Studies simulating propagation through networks have in fact identified network position characteristics that are associated with vulnerability to STIs.6 7 8 However, due to the difficulty of collecting sexual network data, few empirical studies have examined how individual STI risk is affected by sexual-network characteristics. Further, no one has yet tested how improving measurement of the probability of exposure through the use of network data affects our understanding of individual STI risk. To address these gaps in the literature, we used population-based sexual network data from a study combining random digit dialling with snowball sampling to identify network structures associated with STI risk and to evaluate the impact of using network data on epidemiological inferences about STI risk.
Methods
Study design
The Bayview Networks Study was a population-based longitudinal study of STI risk factors among African–American adolescents that took place in San Francisco from 2000 to 2002 and that combined random digit dialing with snowball sampling to generate population-based sexual network data (fig 1).9 10 11 Two waves of data collection were conducted, 12 months apart. At baseline in 2000, a household sample (“index” participants) of 348 African–American adolescents 14–19 years old living in the mixed-income, historically African–American Bayview Hunter’s Point neighbourhood of San Francisco was recruited using random digit dialling, as described in detail previously.10 In 1999, there were approximately 1700 African–American 14–19-year-olds residing in Bayview Hunter’s Point according to the 2000 US census;12 therefore, the 348 baseline index participants comprised roughly 20% of the target population.
The household sample was then used to seed snowball sampling of the adolescents’ social and sexual networks. Sexually active index participants (n = 168, 48%) were asked to name up to two of their social friends (“friends”) and up to six of their sex partners within the past 3 months. Friends who were sexually active were also asked to name their recent sex partners. Named friends and partners were actively recruited by study staff. Recruited sex partners were then asked to name their sex partners within the past 3 months, who in turn were also asked to name their sex partners. These last sex partners, however, were not recruited. At follow-up, a year later, the same index participants and baseline friends were reinterviewed, and their sexual networks were enumerated and interviewed in the same manner as at baseline. While the study was not limited to heterosexual partnerships, only heterosexual partnerships were reported.
All participants had to be at least 14 years old. Informed consent was obtained directly for those 18 and older, while guardian consent with participant assent was obtained for those younger than 18. Participants were asked for permission to contact friends and sex partners they named. Friends and sex partners who were contacted were not told that they had been named by someone. Participants were reimbursed $25 for completing the interview, and $10 for providing a biological sample. All study procedures received human subjects ethics approval from the University of California, San Francisco and the Johns Hopkins Medical Institutions.
Out of 348 baseline index participants, 96% were interviewed at follow-up. Of named friends, 177 (71%) enrolled at baseline and 95% of those participated at follow-up. Over both waves, 49% (770/1581) of named sex partners were enrolled. Non-interviewed partners were less likely to be perceived to be a main partner (47% vs 67%, p<0.01) and more likely to be perceived to have other partners besides the referring partner (33% vs 26%, p<0.01). This information, which suggests that non-interviewed partners were not missing at random, was used in the creation of our exposure variable.
Exposure variable
Individuals were grouped into sexual network components (groups of individuals connected directly or indirectly to each other) based on the sexual partnerships reported between all unique individuals, including those interviewed as well as those named but never interviewed. Two assumptions were made in this process. First, a partnership was considered to exist between two individuals as long as one partner reported it, even if the other partner was interviewed and did not report it, based on the assumption that participants were more likely to omit relationships than to invent them. Second, exact relationship timing was not taken into account, based on evidence that short-term cumulative network measures are more relevant for disease transmission than instantaneous measures.7
We used social network analysis methods13 to calculate sexual network variables likely to be associated with STI risk based on previous studies.6 7 8 14 15 16 We included the following measures of network centrality: degree, two-reach centrality, information centrality, eigenvector centrality and betweenness centrality. We also included the following measures of network structure: component size, density, diameter and centralisation. All network ties were symmetrised reflecting our assumption that, regardless of who reported it, a sexual relationship enabled transmission of an STI in either direction.
Exploratory analyses revealed that only four of these network variables showed statistically significant relationships with STI status: degree (number of partners), two-reach centrality (number of individuals two steps away), component size (number of individuals in a component) and component diameter (longest shortest path between any two component members). Because these variables are strongly related to each other, in order to model their independent and interacting effects we created a composite network position variable based on the combination of degree and two-reach centrality. We focused on degree and two-reach centrality because together they drive component size and diameter. We dichotomised degree (1 vs ⩾2) and two-reach centrality (0 vs ⩾1) to reflect the functional form of the bivariate relationships observed between these two variables and STI risk. This also had the advantage of making the exposure variable less sensitive to missing network data. The possible combinations of these two dichotomous variables created four network positions that simultaneously captured component size and diameter. Further, to take into account the fact that partners who were not interviewed constituted missing data that could not be assumed to be missing at random (see above), we divided those with degree = 1 into two separate positions based on whether their one partner was interviewed. As shown in fig 2, the resulting positions can be described as: (1) members of confirmed dyads (degree = 1 and two-reach centrality = 0 and partner was interviewed), (2) members of unconfirmed dyads (degree = 1 and two-reach centrality = 0 and partner was not interviewed), (3) periphery of non-dyadic components (degree = 1 and two-reach centrality⩾1), (4) centre of star components (degree⩾2 and two-reach centrality = 0) and (5) interior of non-star components (degree⩾2 and two-reach centrality⩾1).
In order to determine how using network data furthers our understanding of individual STI risk, we created a comparable individual-data-based network position variable based solely on data provided by individuals about themselves, that is, ignoring information about who was connected to whom. This individual-based network position variable was based on the individual-level equivalent of degree, which is self-reported number of partners, and the individual-level equivalent of two-reach centrality, which is perceived partner concurrency. Both variables were dichotomised as for degree and two-reach centrality, and the individual-based network position variable was created from the combination of these two dichotomous individual-level variables analogously to the network-based position variable. To replicate the second position, members of unconfirmed dyads, we created a category for individuals with one partner who did not know if their partner had other partners. The resultant five individual-based network positions were each equivalent to the five network-based network positions.
Outcome variable
The dependent variable used in the analysis was current infection with chlamydia and/or gonorrhoea or treatment for these infections within the few weeks preceding the interview. Chlamydia and gonorrhoea were selected as the outcomes of interest because they are the most common STIs that reflect recent behaviours and are therefore ideal outcomes for the study of how sexual networks impact STI acquisition. Current infection with chlamydia and/or gonorrhoea was measured with a ligase chain reaction (LCR) nucleic acid amplification test (LCx Probe System; Abbott Laboratories, Abbott Park, Illinois.) on self-collected vaginal swabs for females and urine samples for males. The LCR test has been found to have a sensitivity of 90–100% and a specificity of 95–100%.17 18
Among sexually active participants of all types (index, friend, sex partner), 77% consented to providing a specimen for STI testing. Individuals who accepted testing were more likely than those who did not to be in more central network positions (p<0.0001). There was no evidence that those who provided samples were more likely to be infected independently of their exposure status.
Recent treatment was assessed during the interview by asking participants when they were last tested and treated for an STI, and for which STI. We included in our analyses those who had recently been treated for gonorrhoea or chlamydia in order to limit outcome misclassification. This contributed eight additional cases to the 96 lab confirmed cases, for a total of 104 cases out of 845 individuals with STI information.
Analytical sample
The sample for this analysis consisted of all recently sexually active participants of any type (index, friend or sex partner) at either wave who either were tested for gonorrhoea (GC) and chlamydia (CT) as part of the study or reported being treated for gonorrhoea or chlamydia within the weeks preceding the interview. The analytical sample included a total of 845 observations from 655 unique individuals, with 397 baseline observations and 448 follow-up observations (table 1). In multivariable models, 54 observations were dropped from the analysis because they were missing condom use, resulting in a sample size of 791. The dropped observations were equally likely to be infected.
Statistical analyses
Bivariate associations were assessed with t tests for continuous variables and χ2 tests for categorical variables. Multivariate analyses were carried out using logistic regression. All p values presented are for two-tailed tests.
There were multiple sets of non-independent observations in our sample: (1) individuals who were interviewed at both waves, (2) individuals in the same network components at a given wave and (3) individuals who were friends with each other. We used logistic models with generalised estimating equations19 to control for these sources of correlation. We could not adjust for all three types of correlation simultaneously because correlated sets were not nested, so we adjusted for each independently. Results from the three adjustments were nearly identical, so we present here only results adjusted for clustering within sexual network components. All multivariate model results are presented with robust standard errors.
Network analysis was conducted in SAS by adapting James Moody’s SPAN modules (SAS Programs for Analyzing Networks, J Moody, Chapel Hill, North Carolina). Pajek (Program for Large Network Analysis, V Batagel; and A Mrvar, Ljubljana, Slovenia) was used to create network images. Statistical analyses were carried out in SAS v.9.1 (SAS Institute, Cary, North Carolina).
Results
Basic demographic and behavioural characteristics of the analytic sample are presented in table 1. At both waves, friends were older than index participants (p = 0.0001) and more likely to be infected with chlamydia (p = 0.06). Compared with both index and friends, sex partners were older (p<0.0001) and less likely to be in school (p<0.005). Prevalence of chlamydia dropped by 40% between baseline and follow-up, possibly due to treatment following baseline testing.
At baseline, among all recently sexually active participants regardless of participant type and STI testing status (n = 518), 781 recent sexual relationships were reported. These comprised 265 separate sexual network components. At follow-up, 576 sexually active participants reported 818 relationships, which formed 295 separate network components. No cycles were observed at either wave (fig 2). Slightly more than half of the components at each wave included only two people (55% at baseline and 59% at follow-up), and only 4% of networks at both waves included 10 or more individuals. The largest component included 28 individuals at baseline and 18 at follow-up. Figure 2 illustrates the location of the sexual network positions and of infected or recently treated individuals in the observed sexual networks.
More than half of all participants at both waves had a degree of one, and roughly half had a two-reach centrality of zero (table 1). At both waves, sex partners were more likely to have a higher degree and two-reach centrality (p<0.0001).
Sexual network position was strongly associated with STI risk (table 2). In models adjusted for age, sex, condom use, participant type and study wave, individuals in the centre of non-dyadic components (positions 4 and 5) had more than six times the odds of infection with chlamydia or gonorrhoea compared with individuals in confirmed dyads (position 1). Individuals on the periphery of non-dyadic components (position 3) had nearly fivefold increased odds of infection, despite having only one partner. Individuals with one partner who was not interviewed (position 2) had more than three times the odds of STIs than individuals with one partner who was interviewed, although this difference was not statistically significant at the p = 0.05 level.
When network position was measured with individual-based data only, associations with network position were underestimated roughly by a factor of 2. Further, individuals in the centres of non-dyadic components (positions 4 and 5) no longer had higher odds of infection than individuals on the periphery of non-dyadic components (position 3) (table 2). While the associations were attenuated in the model using only individual data, they were also more precise because of the larger number of individuals in the reference network position. Associations of STI risk with other variables (age, sex, condom use) were essentially unchanged.
We carried out several sensitivity analyses. Our inferences were unchanged if we assumed observations were clustered over time or within friendships rather than within sexual network components, if we restricted the analysis to the random digit dialling sample at both waves, if we considered as infected only individuals who had lab-confirmed positive results as well as if we limited the analysis to chlamydia-infected individuals only.
Discussion
We set out to empirically investigate the relationship between sexual network position and STI risk, as well as to determine the added value of using network data in addition to individual-based data. We found that sexual network position defined as the interaction between dichotomised degree and two-reach centrality (essentially non-monogamy and partner concurrency) was strongly associated with acquisition of bacterial STIs, controlling for age, gender and condom use. Specifically, individuals in the interior of non-dyadic components were more than six times more likely to be infected with gonorrhoea and/or chlamydia than individuals in confirmed dyads, regardless of whether they were in the centre of star components (position 4) or the interior of non-star components (position 5). Additionally, we found that individuals with only one partner had a substantially elevated risk of being infected if they were on the periphery of a non-dyadic component and therefore connected to a partner who had at least one other partner. We also observed that two-reach centrality did not increase risk among those with more than one partner, as evidenced by the absence of a difference between the odds ratios for positions 4 and 5. Finally, we found that ignoring network linkages underestimated associations by a factor of two and failed to distinguish between risk on the periphery and in the interior of components.
A major limitation of our study is the issue of missing network data. Only 49% of named partners were interviewed due to the difficulty of locating partners and the fact that we conducted only two rounds of snowball sampling. Furthermore, it is likely that some partners were never named because of poor recall or reluctance to disclose partnerships. All of these factors will have led to an underestimation of network connectivity. To address this issue, we used a network position variable designed to minimise sensitivity to missing network data by using dichotomised degree and two-reach centrality and creating a separate network position for individuals with one partner who was not interviewed. Misclassification of this network position variable would primarily occur only if participants failed to name partners. Nonetheless, if misclassification did occur, assuming it was non-differential with regards to infection status, it would only dilute, as opposed to inflate, an association between network connectivity and STI risk. Furthermore, while empirical network studies will always be limited by the inability to fully measure underlying network structure, identification of individuals at elevated risk will ultimately have to be based on empirical observations. Therefore, empirical studies such as this one are precisely what are needed to identify the empirically observable network characteristics associated with increased STI risk.
Another limitation is that we did not conduct interviews with all network members at the same time, but instead as they were identified and located. Given that we used a cumulative network measure, if infected partners were harder to locate, time to find partners might confound the observed association between network structure and STI status. However, we found that time to find partners was independent of the nominating individuals’ or partners’ infection status (51.8 days for dyads involving an infected person vs 48.9 days for dyads with no infected persons on average, p = 0.6).
We used as our outcome infection with either chlamydia and/or gonorrhoea to maximise power. There is evidence, however, that chlamydia transmission may not require the same degree of network connectivity as gonorrhoea due to the higher proportion of chlamydia cases that persist because they are asymptomatic and therefore not treated.20 However, when we replicated our analyses using chlamydia infection only, we found slightly stronger associations with network position. We were unable to carry out multivariate analyses for gonorrhoea alone because of the small number of gonorrhoea cases (26 cases including both waves), however the prevalence of gonorrhoea increased in a dose–response fashion in the five network positions in a manner similar to that seen for chlamydia. We therefore found no evidence that our results depended on which STI we examined.
Our results have several implications for the study and prevention of STI risk. First, we have developed a new network typology that maps strongly onto STI risk and does not necessitate full sociometric network data. Given the time, money and effort needed to collect sociometric data, this network typology may prove useful in future studies of STI risk as well as in interventions aimed at reducing STI incidence through changing network structure.
Second, the variation in risk that we observed among individuals with one partner highlights the need to move beyond number of partners as the main indicator of the probability that an individual will be exposed to an STI. Individuals with one partner can be on the periphery of larger sexual networks and therefore at substantially elevated risk. Our results thereby challenge the notion that the number of partners alone can adequately capture the probability of exposure to STIs and suggest the need for attention to be paid to the sexual network context that places individuals at risk for STIs.
Third, our finding that associations between network position and STIs were twofold underestimated using individual-based data compared with network-based data suggests that epidemiological inferences from studies that rely on individuals’ reports about their partners to measure that sexual network context should be viewed with caution.
Altogether, our results support a shift toward studying and preventing STIs and other infectious disease transmission in a manner that more explicitly takes into account the fact that incidence depends on prevalence21 22 and that therefore takes seriously the need to measure, study and intervene on sexual networks.
Key messages
Sexual network structure is a known determinant of STI propagation, yet much is still unknown about the relationship between network position and STI risk.
This study demonstrates that STI risk is strongly linked to sexual-network position and that the number of partners is an incomplete measure of risk.
Individuals’ reports of partner behaviours are not adequate to accurately capture the probability of exposure to an infected partner.
Future studies should focus on more completely capturing the probability that a person will be exposed to an STI by measuring sexual-network context.
Acknowledgments
The authors thank P Matson and J Jennings for their thoughtful comments.
REFERENCES
Footnotes
Funding This work was supported by grants from the National Institute for Allergy and Infectious Diseases and the National Institute on Drug Abuse.
Competing interests None.
Ethics approval Ethics approval was provided by University of California, San Francisco Johns Hopkins Medical Institutions.
Patient consent Obtained.
Contributors: CMF designed and conducted the analyses and wrote the manuscript; BB led data collection in the field; SQM was responsible for network data management and assisted with network analysis; NSP helped design and oversee data collection; TAG advised on the analysis and interpretation, and helped draft the manuscript; JME conceived and conducted the Bayview Networks Study and helped design the analysis and draft the manuscript.
Provenance and Peer review Not commissioned; externally peer reviewed.
Linked Articles
- Whistlestop tour