From the University of Washington, Seattle, Washington
Correspondence: Martina Morris, PhD, University of Washington, Box 353340, 1959 N. East Pacific St., Seattle, WA 98195. E-mail: firstname.lastname@example.org
Received for publication June 28, 2001 and accepted July 7, 2001.
THE LAST DECADE has witnessed a paradigm shift in the field of sexually transmitted infection (STI) epidemiology. Our analytic focus has extended beyond the knowledge, attitudes, and behavior of individuals, to incorporate the structure and dynamics of partnership networks within which behavior becomes exposure. This has changed the way we work, from the way we design studies and analyze data, to our understanding of the distribution of risk and the dynamics of transmission, and finally to our approach to developing strategies for prevention. The article in this issue by Koumans and coworkers 1 is a good example of this new approach. Koumans et al find that concurrent partnerships are strongly implicated in the transmission of syphilis in the United States, thus perhaps helping to explain a part of the puzzle of syphilis persistence in low rate settings. Their findings complement a rapidly growing body of literature on concurrent partnerships. Together, these studies now provide a compelling picture of the central importance of concurrency in the transmission of infections.
Concurrent partnerships are partnerships that overlap in time, rather than follow one another sequentially and disjointedly. Their potential relevance for STI was first suggested at the beginning of the 1990s in the context of the HIV epidemic in sub-Saharan Africa. 2,3 Further simulation studies have shown that concurrent partnerships can strongly amplify the initial spread of infection, 4 even more strongly when the aggregate mixing pattern is highly assortative or disassortative. 5 Transmission is amplified because concurrent partnerships link individuals together to create large connected components—if you have more than one partner, then your partner may have more than one partner, and so on. Such connected components function like a well designed road network enabling a pathogen to travel rapidly and efficiently to many destinations.
Early simulation findings, and the intuition behind them, played a major role in placing the question of concurrent partnerships on the research agenda, 6,7 and it is remarkable how quickly that research has produced results. On the empirical side, a number of studies have now documented the prevalence of concurrent partnerships in populations with high rates of circulating STI, both in clinic settings, 8,9 and in community-based settings. 6,10–13 We also now know that concurrency is associated with higher rates of transmission. From the work of Potterat and colleagues, 14 we have the estimate that the adjusted odds of transmitting chlamydia are 3.2 times higher for index cases with concurrent partnerships. This is almost exactly the same as the estimate obtained by Koumans et al in this study, where the adjusted odds of being a syphilis transmitter are 3.1 for cases with concurrent partnerships. On the theoretical side, work has been done to develop appropriate methodological tools for epidemiologic network modeling, 15–17 and Bauch and Rand 18 have now shown that concurrency increases R0, the basic reproductive rate of an epidemic and the fundamental measure of epidemic potential.
The units of a transmission network are persons and partnerships, which in turn are cumulated up to form larger structures. The reason concurrency is so important is that it is the basic mechanism by which a partnership network becomes connected at a particular moment in time. Without it, we have only dyads (pairs) and isolates that trap the pathogen and prevent it from spreading. With it, we create triples (as Koumans et al use here) and larger components of all kinds. While the formation and dissolution of sequential partnerships over time also provides a type of connection between partnerships, this is a slower process, and the sequencing provides protection for the earlier partners.
We have gained a great deal of theoretical insight and practical knowledge from the network approach to STI epidemiology, but we are also now more aware of the difficulties it entails. Working with units of analysis beyond individuals (pairs, triples, etc.) creates a serious challenge for empirical work. Our traditional methods of data collection and analysis are deeply rooted in the assumption that the isolated, independent individual is the unit of analysis. Network analysis requires finding ways to measure properties that cannot be readily observed at the individual level, and developing appropriate statistical methods for dealing with the dependence and interdependence of units within a network. It also involves subtle changes in perspective that set traps for the unwary.
One such a trap is the use of relative risk analysis, the workhorse of empirical epidemiology, for estimating the “effect of concurrency.” It seems the most natural thing in the world to predict the disease status of an index case as a function of concurrent partnerships, controlling for the number of partners in order to see if concurrency has an independent effect. But, in general, this will be wrong. It is relatively easy to see this in the simple case of monogamous women whose male partners have other partners. The men’s concurrency here puts their monogamous female partners at risk, especially when the concurrent partners are sex workers. 19 A traditional logistic regression would not attribute the monogamous women’s infections to concurrency, however, because these women do not themselves have concurrent partners. One might be tempted to try to avoid this problem by restricting the sample to persons with at least two partners. But this does not solve the problem either, because the basic point is that concurrency creates a risk for the partner, not the index case.
To see this, consider the two scenarios in Figure 1. In the first case, the survey respondent (A) has two partners (B and C) concurrently at time 2. In the second case, A has these two partners sequentially without overlap. Logistic regression would typically be used to compare infection prevalence in the respondents (As) from scenario 1 and 2, estimating the effect of concurrency as the relative risk of infection between these two groups. In both scenarios, A is exposed to any infection that B and C may have. So there is no reason to expect that the concurrency in scenario 2 would increase the probability of infection for A, over and above the risk of having two partners. The only increase in risk is for partner B, who is exposed indirectly to C due to concurrency in scenario 1, but not in scenario 2. (We have ignored the issue of duration here to simplify exposition, but it will introduce another level of complexity.) As before, concurrency is not a risk for the index case A, but for A’s partners, B and C. That is why a logistic regression of individual infection on individual concurrency is simply not appropriate: it is measuring the wrong thing. Moreover, concurrency is not a risk for both of A’s partners equally. The greatest risk is for partner B, the partner who placed earlier in the sequence (and for any other partners B might have at time 2). Partner C would be indirectly exposed to partner B in both scenarios, so his or her risk would only increase if B took on additional partners during time 2. We might also expect that the impact of concurrency will vary inversely with the duration of infection. The shorter the duration, the faster one has to find another partner to continue the chain of transmission. With concurrency, that partner is already there. Identifying the effect of concurrency thus clearly requires careful thinking about how and what to measure.
Koumans et al have avoided the error of trying to estimate the effect of concurrency on the index case A. Instead they have chosen to measure whether A is more likely to be an infection transmitter in scenario 1 as compared with scenario 2. This innovative research design parallels that used in the Potterat et al 14 study of chlamydia. Both studies have measured the right thing at the right level: whether A’s partners are more likely to be at risk of infection if A has them concurrently rather than sequentially. Because they do not distinguish between partners that are earlier or later in the sequence, some misspecification of the measurement model remains. Their approach estimates the average impact of concurrency for earlier and later partners.
Note that what makes it possible for Koumans, Potterat, and their colleages to conduct the analysis of A as an infection transmitter rather than receiver is the link-tracing sample design. They have enrolled the partners of A, and can therefore establish the partners’ disease status. Such link-tracing designs are expensive, invasive, and generally unfeasible outside of the public health contact-tracing context. This points to the question of how best to collect data on networks—a question that is an active area of theoretical research. Koumans and associates employ a variant of a “partial network” design using link tracing to create a type of snowball sample. 20 Contact tracing in this particular context stops when a partner is found to be uninfected. This creates a form of sample selection bias if all enrolled persons are analyzed in a pooled sample (which is not done here). As an alternative, the frequency of concurrent partnerships can be established using more standard random survey sampling methods for “local networks,” by asking survey respondents the beginning and ending dates of (some sample of) their partnerships. Local network designs do not enroll the partners, however, so their disease status can not be verified. As a result, survey respondents in these studies can only be identified as infection receivers, not transmitters. Local network designs are thus generally better suited for estimating the prevalence of concurrency rather than its effects on relative risk. The larger point, though, is that both sampling strategies produce an incomplete picture of the network. There is much work that needs to be done before we understand the many ways in which the missing data produced by different sample designs influence our ability to estimate network effects.
Given the constraints of our current analytic tool kit for networks, what Koumans and her colleagues accomplish in this study is all the more impressive. Their contribution is not to verify that concurrency increases the spread of infection—that we already knew from simulation. Their contribution is instead to show that concurrency is prevalent in the populations where syphilis has managed to maintain a presence, and that it is associated with higher rates of transmission. Koumans et al have used the unique features of their data to tease out the interesting, and testable, hypothesis that concurrency may play a central role in maintaining persistent low rate endemic syphilis in the United States. If it proves to be correct, then Koumans and her colleagues have identified concurrent partnerships as an important strategic target of intervention for syphilis eradication.
Prevention will return us to the question of individual behavior, but with an explicitly relational perspective. We need to understand the range of different types of concurrent partnerships, the social, economic and cultural conditions under which they are established, the meaning attached to such relational patterns, and the variation in partner-specific behavior. This relational perspective will help us to systematically take into account the interpersonal context of risk behavior, and to better identify the constraints and opportunities for behavioral change.
1. Koumans EH, Farley TA, Gibson JJ, et al. Characteristics of persons with syphilis in areas of persisting syphilis in the United States: sustained transmission associated with concurrent partnerships. Sex Transm Dis 2001; 28: 497–503.
2. Watts CH, May RM. the influence of concurrent partnerships on the dynamics of HIV/AIDS. Math Biosci 1992; 108: 89–104.
3. Hudson C. Concurrent partnerships could cause AIDS epidemics. Int J STD AIDS 1993; 4: 349–353.
4. Morris M, Kretzschmar M. Concurrent partnerships and the spread of HIV. AIDS 1997; 11: 641–8.
5. Morris M, Kretzschmar M. Concurrent partnerships and transmission dynamics in networks. Soc Net 1995; 17: 299–318.
6. Hudson C. AIDS in rural Africa: a paradigm for HIV-1 prevention. Int J STD AIDS 1996; 7: 236–243.
7. Garnett G, Johnson A. Coining a new term in epidemiology: concurrency and HIV. AIDS 1997; 11: 681–683.
8. Howard MM, et al. Patterns of sexual partnerships among adolescent females. J Adolesc Health 1999; 24: 300–303.
9. Rosenberg M, Gurvey J, Adler N, Dunlop M, Ellen J. Concurrent sex partners and risk for sexually transmitted diseases among adolescents. Sex Transm Dis 1999; 26: 208–212.
10. Colvin M, SS A-K, Connolly C, Hoosen A, Ntuli N. HIV infection and asymptomatic sexually transmitted infections in a rural South African community. Int J STD AIDS 1998; 9: 548–550.
11. Ford K, Norris A. Patterns of union formation among urban minority youth in the United States. Arch Sex Behav 2000; 29: 177–187.
12. Morris M, Kretzschmar M. A micro-simulation study of the effect of concurrent partnerships on HIV spread in Uganda. Math Popul Stud 2000; 8.
13. Rothenberg R, Long D, Sterk C, et al. The Atlanta Urban Networks Study: a blueprint for endemic transmission. AIDS 2000; 14: 2191–2200.
14. Potterat J, Zimmerman-Rogers H, Muth S, et al. Chlamydia transmission: concurrency, reproduction number, and the epidemic trajectory. Am J Epidemiol 1999; 150: 1331–1339.
15. Kretzschmar M, Morris M. Measures of concurrency in networks and the spread of infectious disease. Math. Biosci 1996; 133: 165–195.
16. Chick S, Adams A, Koopman J. Analysis and simulation of a stochastic, discrete-individual model of STD transmission with partnership concurrency. Math Biosci 2000; 166: 45–68.
17. Ferguson N, Garnett G. More realistic models of sexually transmitted disease transmission dynamics: sexual partnership networks, pair models, and moment closure. Sex Transm Dis 2000; 27: 600–609.
18. Bauch C, Rand D. A moment closure model for sexually transmitted disease transmission through a concurrent partnership network. Proc R Soc Lond B Biol Sci 2000; 267: 2019—2027.
19. Morris M, Podhisita C, Wawer M, Handcock M. Bridge populations in the spread of HIV/AIDS in Thailand. AIDS 1996; 11: 1265–1271.
20. Morris M. Sexual Networks and HIV. AIDS 1997; 11 (suppl A): S209–S216.