Temporal patterns in sexual partnerships play an important role in HIV and sexually transmitted infection (STI) transmission dynamics. Concurrency in particular has received considerable attention1–6 because it is hypothesized to contribute to rapid, extensive spread of HIV and other STIs.7–10 However, not all patterns of concurrency confer the same level of transmission risk,11–14 and other partnership patterns beyond concurrency can influence transmission. Sexual partnership duration,15–17 gap length between partnerships,15,16,18–22 and overlap length across concurrent partnerships15,16,23 are important determinants of the rate and reach of HIV/STI transmission and capture temporal patterns in partnerships that are missed by standard measures of concurrency.11 Measuring these patterns and understanding how they impact HIV/STI transmission are critical for developing targeted interventions to prevent transmission.
Estimating temporal patterns requires a clear definition of sexual partnerships and rules for its operationalization. Sexual partnerships are commonly defined as the time between date of first sex (partnership start) and date of last sex (partnership end).15–20,22 The date of partnership end can be difficult to establish if a partnership is ongoing at the time of data collection because the respondent can only provide the date of most recent sex (i.e., the date she/he last had sex with the partner before the interview), not the date of true last sex. To address this issue, some studies assess whether participants believe that their partnerships are ongoing and incorporate this information into analyses.
To date, there have been 2 main approaches for measuring partnership duration. One approach, which we call the “first-generation” approach, typically does not gather information about partnership ongoing status and thus assumes all partnerships have ended at the time of data collection (i.e., assumes that most recent sex is truly last sex). Under this approach, partnership duration is estimated by taking the difference between the dates of first and most recent sex.15,16,24 When partnerships have truly ended, this approach estimates partnership duration (Fig. 1A, C). When partnerships are ongoing, this approach underestimates partnership duration (Fig. 1B, D)17 and concurrency (Fig. 1F, H) because it misclassifies ongoing partnerships as ended rather than right-censoring them. Nevertheless, this approach does evaluate partnerships as they are at the time of the interview without introducing error/uncertainty due to respondents predicting future sex.
Another approach, which we call the “second-generation” approach, assesses partnership ongoing status and incorporates this information by taking the difference between dates of first and most recent sex if the partnership has reportedly ended, and first sex and interview if the partnership is reportedly ongoing. Partnership durations are then estimated using Kaplan-Meier methods, and reportedly ongoing partnerships are right-censored (Fig. 1C, D).17 However, if respondents incorrectly determine whether partnerships are ongoing, this approach biases partnership duration and concurrency (Fig. 1B, C, F, G).11,25,26
Given these concerns, we explore an alternative “hybrid” approach where respondents still classify partnerships as ongoing/ended, but only those reportedly ongoing partnerships in which sex occurred recently are right-censored (where “recently” is defined as the past 3 or 6 months).
We examine how estimates of partnership duration, gap length, and overlap length vary under the first-generation, second-generation, and hybrid approaches in a cohort of sexually active adolescent girls living in rural South Africa. In addition, we estimate the lifetime prevalence of concurrency and assess the potential for rapid HIV/STI transmission in relation to estimated gap lengths.
Study Design, Setting, and Population
This secondary analysis uses baseline data from the HIV Preventions Trial Network (HPTN) 068 study, a randomized controlled trial of cash transfers for HIV prevention among 2533 unmarried school girls aged 13 to 20 years. Baseline data were collected between March 2011 and December 2012 from young women living in rural Mpumalanga Province, South Africa, in households situated in the Agincourt Health and Sociodemographic Surveillance System.27 This study focuses on a subcohort of 654 adolescent girls who reported ever having sex at baseline and provided information on the date of first sex, date of most recent sex, and partnership status (ongoing or not) for at least 1 sexual partnership.
Ethics approval for the study was obtained from the University of North Carolina Institutional Review Board, the University of the Witwatersrand Human Subjects Ethics Committee, and the Mpumalanga Departments of Health and Education. Assent and informed consent were obtained from the girls and their parent/legal guardian, respectively, at study enrollment.
Adolescent girls were interviewed using audio computer-assisted self-interview at study enrollment about the exact date of first sex (What was the month, day, year that you first had sex with [partner]?), date of most recent sex (What was the month, day, and year that you last had sex with [partner]?), and partnership status (Is this partnership ongoing or ended?) at the time of interview for their 3 most recent sexual partners. In situations where one partnership ended on the same day that another partnership started, we assumed that the partnerships were separated by a gap of 0.5 days. Partnerships with implausible start or end dates (e.g., date of first sex was after the date of most recent sex, date of most recent sex was before the date of first sex) or missing start or end dates were excluded from the analysis (with the exception of reportedly ongoing partnerships with missing end dates under the second-generation approach, where we assumed the date of the interview was equal to the date of most recent sex). We did not limit the period during which girls could report on their sexual partnerships, avoiding the length time bias that can be introduced when sampling windows (such as allowing participants to only report on partners in the past year) are used.17
We used information on date of first sex, date of most recent sex, and partnership status (ongoing or ended) to estimate median partnership duration, gap length between partnerships, and overlap length across concurrent partnerships. We used 4 measurement approaches that implemented different rules in determining when partnerships had ended (Table 1). We first calculated crude partnership durations, gap lengths, and overlap lengths under each approach before estimating the distributions of these measures using Kaplan-Meier with a robust variance estimator to account for correlation due to girls reporting multiple partnerships.28 Across all approaches, crude values that equaled zero were reassigned a value of 0.5 days to enable their inclusion in the Kaplan-Meier distribution. All analyses were performed using SAS v.9.3 (SAS Institute, Cary, NC).
Defining and Estimating Partnership Parameters: First-Generation Approach. Under the first-generation approach, we assumed that all partnerships ended at the time of most recent sex regardless of reported ongoing status. We calculated crude partnership duration by taking the difference between the date of first sex and most recent sex (Fig. 1A–D). Among sets of nonconcurrent partnerships, we calculated crude gap length by taking the difference between the date of most recent sex for the earlier partner (MR1) and the date of first sex for the newer partner (F2) (where “earlier” vs. “newer” is determined by the date of first sex; Fig. 1E–H). Among concurrent partnerships, we calculated crude overlap length following methods outlined in Powers et al.15 and Mercer et al.20 If the partnerships were partially contained (i.e., 1 partnership partially overlapped another), we calculated overlap length by taking the difference between the date of first sex for the newer partner (F2) and the date of most recent sex for the earlier partner (MR1). If one partnership was completely contained in the other, we calculated overlap length by taking the difference between the dates of first (F2) and most recent sex (MR2) for the fully contained partnership. We then estimated the distribution of these crude measures—partnership duration, gap length, and overlap length—using a Kaplan-Meier approach assuming no censoring.
Defining and Estimating Partnership Parameters: Second-Generation Approach. Under the second-generation approach, we assumed that if partnerships were reportedly ongoing, sex would continue up to and past the date of the interview. Therefore, we reassigned the date of most recent sex-to the date of the interview in these partnerships. Importantly, this reassignment changed concurrent partnerships that were completely contained under the first-generation approach to partially contained under the second-generation approach, if the newer partnership was reportedly ongoing. It also changed concurrent partnerships that were partially contained under the first-generation approach to completely contain under the second-generation approach, if the earlier partnership was reportedly ongoing but the newer partnership reportedly ended. Finally, partnerships that were separated by a gap under the first-generation approach became concurrent under the second-generation approach, if the earlier partnership was reportedly ongoing.
Under this approach, both partnership duration and overlap length could be ongoing. Overlaps were considered ongoing if both concurrent partnerships were reportedly ongoing. We treated ongoing partnerships and overlaps as right-censored when estimating their distribution using a Kaplan-Meier estimator. For partnerships or overlaps that reportedly ended, and for all gaps, we calculated crude estimates according to the first-generation approach and treated these measures as not censored.
Defining and Estimating Partnership Parameters: Hybrid-6 Month Approach. We followed the second-generation approach for estimating partnership durations, gap lengths, and overlap lengths, but instead of basing censoring determination solely on self-reported partnership status, we considered partnerships to be right-censored (i.e., ongoing) only if the girl reported (1) that the partnership was ongoing and (2) that she had sex with her partner within the last 6 months (i.e., time from most recent sex to interview was ≤181 days). If a partnership was reportedly ongoing but there was no sexual activity in the last 6 months, we administratively ended the partnership at the time of most recent sex and treated it as not censored in the analysis.
Defining and Estimating Partnership Parameters: Hybrid-3 Month Approach. We took the same approach described under the hybrid-6 month approach, except we used a cutoff of 3 months to administratively end partnerships.
Assessing Implications for HIV and STI Transmission: Concurrency and Short Gap Lengths That Can Facilitate Transmission. Lifetime concurrency status was determined using self-reported dates of first and most recent sex for the 3 most recent sexual partnerships and was defined as having any partnership set where the date of first sex for the newer partner occurred before the date of most recent sex for the earlier partner.
Serially monogamous partnerships can facilitate HIV/STI transmission if the gap length between partnerships is shorter than the remaining, highly infectious early HIV infection period or shorter than the remaining infectious period for an STI.18,20 We explored the potential for transmission by short gap lengths between serially monogamous partnerships by estimating the percentage of partnership gaps that were shorter than 3, 6, 9, 12, and 24 months, choosing these cut points to approximate infectious periods of common STIs and the early HIV infection period. We also generated Kaplan-Meier survival curves of the distribution of gap lengths across the 4 approaches to examine their relationships to the infectious periods of selected STIs and early HIV infection.
Of the 2533 adolescent girls eligible for the parent study, we excluded 1840 girls who reported no history of sex, 32 sexually experienced girls with missing dates of first/most recent sex for their only reported sexual partnership, and 7 sexually experienced girls who reported a future date of first sex for their only reported sexual partnership. The remaining 654 girls were eligible for analysis and contributed a total of 1066 sexual partnerships.
Mean age was 17 years, mean age at first sex was 14.6, and 85% of girls reported 3 or less lifetime sexual partners (Table 2). Sixty-five percent of girls reported that their most recent partner was more than 2 years older and 41% reported not using a condom at last sex. Six percent of girls were HIV positive (n = 40) and 13% had herpes simplex virus type 2 (HSV-2; n = 84).
Partnership Duration, Gap Length Between Partnerships, Overlap Length Across Partnerships
Median partnership duration ranged from 368 days (95% confidence interval [CI], 338–424) to 1024 days (95% CI, 810–1531) under the first- and second-generation approaches (Table 3), respectively. Both hybrid approaches produced intermediate estimates (387 and 595 days), with the shorter estimate under the hybrid-3 month approach. Under the hybrid-6 month and hybrid 3-month approaches, we administratively ended 12% (n = 125) and 19% (n = 200) of all reportedly ongoing partnerships, respectively.
Estimated gap lengths were relatively stable across approaches, with median values ranging from 143 days (95% CI, 96–194) under the first-generation approach to 185 days (95% CI, 137–262) under the second-generation approach (Table 3). Both hybrid approaches were within this narrow range.
Median overlap length across concurrent partnerships ranged from 168 days (95% CI, 101–237) under the first-generation approach to 409 days (95% CI, 274–919) under the second-generation approach (Table 3). Estimates from both hybrid approaches were intermediate, with a shorter overlap estimate from the 3-month approach (185 days) than the 6-month approach (240 days).
Implications for HIV/STI Transmission
The lifetime prevalence of concurrency was stable across the 4 approaches, ranging from 28% to 33% (Table 4). The proportion of gaps shorter than 3, 6, 9, 12, and 24 months was also consistent across approaches, ranging from 32% to 39%, 50% to 55%, 61% to 67%, 69% to 73%, and 88% to 90%, respectively (Table 5). The distribution of estimated gap lengths suggests considerable potential for short gap lengths between serially monogamous partnerships (Fig. 2) to facilitate transmission of HIV and several common STIs.
We estimated sexual partnership durations, gap lengths between partnerships, and overlap lengths across concurrent partnerships among sexually experienced adolescent girls living in rural South Africa. Adolescent girls in this context are at extremely high risk for HIV acquisition29 and their partnership patterns may have important implications for disease transmission. These girls reported long sexual partnerships, which can reduce the risk of HIV/STI acquisition when partnerships are monogamous and both partners are uninfected. However, the girls also reported high levels of concurrency and short gaps between partnerships, which can facilitate transmission.18,20 Gap length may be especially important for STIs with short infectious periods like gonorrhea, which can be sustained in a population when only a small group of individuals exhibit short gaps and medium partnership lengths.18
We relied on cross-sectional data which poses challenges for estimating ongoing partnership parameters, including partnership duration and the prevalence and length of concurrency. Specifically, such data force us to rely on girls' ability to predict their future sexual behavior rather than directly measuring this behavior longitudinally. Studies that longitudinally collect partnership data do not require participants to predict future behavior, but they do not fully eliminate the problem of censoring due to ongoing partnerships at the end of a cohort study. An optimal approach for addressing ongoing partnerships is still needed for both cross-sectional and longitudinal data.
Estimates of these parameters are important for HIV/STI transmission modeling and for understanding the potential impact of prevention programs. However, there are considerable challenges in how these parameters are estimated, and a validated, optimal approach has not been established. Kaplan-Meier approaches that rely on self-reported partnership ongoing status (i.e., the second-generation approach) have been recommended for partnership duration estimation,15,17,18 but neither the reliability of such self-reports nor the effects of different assumptions about their reliability have been assessed. We addressed the latter issue and found that estimates of partnership duration and overlap length varied considerably across approaches (e.g., the hybrid-3 month approach used a more “stringent” definition of ongoing partnership than did the hybrid-6 month or second-generation approaches, resulting in shorter durations).
Notably, we observed many cases where girls reported that partnerships were ongoing even when more than 6 months had passed since most recent sex. This finding may suggest that adolescents have difficulty determining when a partnership is “ongoing” and that analysis approaches reliant on self-report for determining censoring may be unreliable. Alternatively, if these girls are in fact accurately forecasting future sexual contact with their partners, these findings could suggest that long breaks in sexual activity are common in this population. Indeed, long partnerships may be interspersed with breaks due to a migrant partner30 or because partners break up and get back together again.24 On-again/off-again partnerships may be particularly common among adolescents who are experimenting sexually with different partners.
Thus, our results not only highlight how sensitive estimates are to how censored status is determined; they also raise fundamental questions about how partnerships are defined for studying HIV/STI transmission. One underappreciated feature of the second-generation approach is that it administratively adds time (equal to the time between most recent sex and interview) to reportedly ongoing partnerships when we explicitly know that sex did not occur during this period. Because many applications of these types of partnership duration estimates assume that within-partnership behaviors are constant across the entire estimated duration, any resulting transmission rate predictions may be exaggerated. In sum, the standard definition of sexual partnerships (time from first sex to last sex), the corresponding second-generation estimation approach (Kaplan-Meier with censoring based on self-reported ongoing status), and common applications of these estimates (e.g., in transmission models) may all be too simplistic to capture important features of sexual partnership patterns in many circumstances.
Based on these observations, we suggest several avenues for future research. First, we advocate that future studies evaluate the extent to which individuals are able to predict their future sexual behavior, and the reliability of alternate approaches (e.g., administrative ending as performed in the “hybrid approach”) to determine partnership end. In addition, we encourage the development of more complex concepts of a “partnership” beyond a single duration with assumed uniformity of behaviors within it. More nuanced approaches could account for temporal changes in coital frequency, partner migration, and on-again/off-again partnerships. Such concepts can be iteratively developed alongside studies that measure more complex partnership patterns and behaviors over time to better understand dynamic partnerships patterns and their implications for transmission. We acknowledge that such studies will be challenging to implement, but our findings suggest that further thought and research into these issues is important.
We note that the hybrid approach used in these illustrative analyses only highlights potential biases of the second-generation approach arising from respondents predicting that partnerships will continue when they in reality will not. We did not consider measurement bias due to girls misreporting ongoing partnerships as ended (Fig. 1B, F), although some bias in this direction may also be likely and could be a valuable topic for future work. In addition, we focused specifically on comparing analytical approaches that differed in their determination of partnership end, although other estimation approaches exist and warrant further investigation. One approach that was not compared, but has been used in at least one modeling study,31 uses the mean age only of reportedly ongoing partnerships to estimate partnership duration. This approach makes a number of assumptions, including (1) that the mean age of ongoing relationships at a randomly selected point in time is equal to the expectation for the mean duration of relationships after completion over a long period of time, and (2) that the distribution of partnership durations is geometric, resulting in left and the right censoring cancelling each other out. Future studies should examine this approach, the plausibility of its assumptions, and its potential benefits and tradeoffs.
In summary, measures of partnership patterns that take into account temporal aspects of partnerships, including partnership duration, gap length, and overlap length, are critical for understanding transmission dynamics and designing effective prevention programs. Standardized measures of concurrency11 can provide some insight into HIV and STI transmission, but they fail to fully capture the rich diversity and multifaceted nature of partnership patterns. More detailed characterizations of partnership patterns and how they influence transmission can help researchers design more effective, targeted interventions. Measuring partnership patterns is challenging, and we demonstrated that estimates are sensitive to methods for assessing and accounting for ongoing partnerships. Additional studies that track partnerships longitudinally and with frequent follow-up intervals, along with additional analytical methods and conceptual frameworks for describing partnership dynamics, are needed to better characterize partnership patterns in the context of infectious disease transmission.
1. Morris M. Barking up the wrong evidence tree. Comment on Lurie & Rosenthal, “Concurrent partnerships as a driver of the HIV epidemic in sub-Saharan Africa? The evidence is limited.”. AIDS Behav 2010; 14: 31–33.
2. Lurie MN, Rosenthal S. The concurrency hypothesis in sub-Saharan Africa: Convincing empirical evidence is still lacking. Response to Mah and Halperin, Epstein, and Morris. AIDS Behav 2010; 14: 34.
3. Epstein H. The mathematics of concurrent partnerships and HIV: A commentary on Lurie and Rosenthal, 2009. AIDS Behav 2010; 14: 29–30.
4. Lurie M, Rosenthal S, Williams B. Concurrency driving the African HIV epidemics: Where is the evidence? Lancet 2009; 374: 1420 author reply 1420-1. doi: 10.1016/S0140-6736(09)61860-2.; 2009.
5. Mah TL, Halperin DT. Concurrent sexual partnerships and the HIV epidemics in Africa: Evidence to move forward. AIDS Behav 2010; 14: 11–16.
6. Mah TL, Halperin DT. The evidence for the role of concurrent partnerships in Africa's HIV EPidemics: A response to Lurie and Rosenthal. AIDS Behav 2010; 14: 25–28.
7. Morris M, Kretzschmar M. A microsimulation study of the effects of concurrent partnerships on the spread of HIV in Uganda. Math Popul Stud 2000; 8: 109–133.
8. Kretzschmar M, Morris M. Measures of concurrency in networks and the spread of infectious disease. Math Biosci 1996; 133: 165–195.
9. Goodreau SM, Cassels S, Kasprzyk D, et al. Concurrent partnerships, acute infection and HIV epidemic dynamics among young adults in Zimbabwe. AIDS Behav 2012; 16: 312–322.
10. Eaton JW, Hallett TB, Garnett GP. Concurrent sexual partnerships and primary HIV infection: A critical interaction. AIDS Behav 2011; 15: 687–692.
11. UNAIDS. Consultation on concurrent sexual partnerships: Recommendations from a meeting of the UNAIDS Reference Group on Estimates, Modelling and Projections. 2009.
12. Gorbach PM, Stoner BP, Aral SO, et al. “It takes a village”: Understanding concurrent sexual partnerships in Seattle, Washington. Sex Transm Dis 2002; 29: 453–462.
13. Lurie MN, Rosenthal S. Concurrent partnerships as a driver of the HIV Epidemic in sub-Saharan Africa? The evidence is limited. AIDS Behav 2010; 14: 17–24.
14. Kretzschmar M, White RG, Caraël M. Concurrency is more complex than it seems. AIDS 2010; 24: 313–315.
15. Powers KA, Hoffman IF, Ghani AC, et al. Sexual partnership patterns in Malawi: Implications for HIV/STI transmission. Sex Transm Dis 2011; 38: 657–666.
16. Foxman B, Newman M, Percha B, et al. Measures of sexual partnerships: Lengths, gaps, overlaps, and sexually transmitted infection. Sex Transm Dis 2006; 33: 209–214.
17. Burington B, Hughes JP, Whittington WL, et al. Estimating duration in partnership studies: Issues, methods and examples. Sex Transm Infect 2010; 86: 84–89.
18. Chen MI, Ghani AC, Edmunds J. Mind the gap: The role of time between sexes with two consecutive partners on the transmission dynamics of gonorrhea. Sex Transm Dis 2008; 35: 435–444.
19. Zhan W, Krasnoselskikh TV, Golovanov S, et al. Gap between consecutive sexual partnerships and sexually transmitted infections among STI clinic patients in St Petersburg, Russia. AIDS Behav 2012; 16: 334–339.
20. Mercer CH, Aicken CR, Tanton C, et al. Serial monogamy and biologic concurrency: Measurement of the gaps between sexual partners to inform targeted strategies. Am J Epidemiol 2013; 178: 249–259.
21. Castillo-Guajardo D, García-Ramos G. Estimates of sexual partnership dynamics: Extending negative and positive gaps to status lengths. J Epidemiol Community Health 2010; 64: 672–677.
22. Kraut-Becher JR, Aral SO. Gap length: An important factor in sexually transmitted disease transmission. Sex Transm Dis 2003; 30: 221–225.
23. Steffenson AE, Pettifor AE, Seage GR 3rd, et al. Concurrent sexual partnerships and human immunodeficiency virus risk among South African youth. Sex Transm Dis 2011; 38: 459–466.
24. Matson PA, Chung SE, Ellen JM. When they break up and get back together: Length of adolescent romantic relationships and partner concurrency. Sex Transm Dis 2012; 39: 281–285.
25. Eaton JW, McGrath N, Newell ML. Unpacking the recommended indicator for concurrent sexual partnerships. AIDS 2012; 26: 1037–1039.
26. Bärnighausen T, Hosegood V, Timaeus IM, et al. The socioeconomic determinants of HIV incidence: Evidence from a longitudinal, population-based study in rural South Africa. AIDS 2007; 21(Suppl 7): S29–S38.
27. Kahn K, Collinson MA, Gómez-Olivé FX, et al. Profile: Agincourt health and socio-demographic surveillance system. Int J Epidemiol 2012; 41: 988–1001.
28. Williams RL. A note on robust variance estimation for cluster-correlated data. Biometrics 2000; 56: 645–646.
29. Pettifor AE, Rees HV, Kleinschmidt I, et al. Young people's sexual health in South Africa: HIV prevalence and sexual behaviors from a nationally representative household survey. AIDS 2005; 19: 1525–1534.
30. Collinson MA. Striving against adversity: The dynamics of migration, health and poverty in rural South Africa. Glob Health Action 2010; 3.
31. Goodreau SM, Carnegie NB, Vittinghoff E, et al. What drives the US and Peruvian HIV epidemics in men who have sex with men (MSM)? PLoS One 2012; 7: e50522.