The ultimate goal of any HIV prevention program is to reduce the number of new infections. Evaluating the impact of HIV/STI interventions on reducing HIV transmission is, therefore, an essential part of overall prevention and control efforts. However, this type of program assessment remains an elusive goal for most AIDS control programs because of the prohibitive costs and methodological difficulties associated with field-based program impact evaluation. As a consequence, the important question of the extent to which program effects may lead to reductions in HIV transmission is usually not answered by intervention programs.

In response to these challenges, Family Health International (FHI) has developed the AVERT computer model, which may offer a useful option to address that question1. The AVERT model can be used to estimate the impact of prevention interventions, such as those that focus on increasing use of condoms, improving treatment of sexually transmitted infections (STIs), or changing sexual behaviors, on the reduction of primary HIV transmissions through sexual intercourse over a given time period.

Various types of computer models of the AIDS epidemic have been devised for various purposes. Many, if not most of them, are relatively complex and some can only be used as research tools by their creators. There is a clear need for less complex models that are accessible to epidemiologists and other public health specialists who have limited time and resources. The AVERT model answers this need in a manner consistent with the availability of local program data, which means that data requirements are considerably less extensive than those for more sophisticated simulation models such as SimulAIDS or iwgAIDS2.

While AVERT is much easier to use than other models, its static nature is not appropriate for modeling long-term scenarios. AVERT cannot be used to explore some of the more complex questions that dynamic simulation models aim to address in the context of hypothesis testing.

This was a conscious strategy to keep the model accessible to non-experts in modeling and to accommodate the types of data available to most intervention programs.

This chapter explains the approach used in AVERT and describes the required input variables. Estimates generated by the model are validated against seroconversion data from a large cohort study and practical examples of its application are presented.

Description Of Avert

AVERT is a relatively simple, static, mathematical model with an interface developed especially for users with little experience in modeling. The model requires relatively few input parameters, all of which may be readily available to programs, and it focuses on population dyads, namely hetero/homosexual couples in union or female sex workers and their clients. Using changes in four major intervention parameters that affect transmission rates, the model calculates how many fewer infections there would be in a defined time period. The parameters are:

• number of partners;
• number of sex acts per partner;
• condom use; and
• STI levels.

Model estimates are usually based on a 1-year time frame.

Developing this relatively simple and easy-to-use model involved an important trade-off in that some of the complexities of the epidemic are not incorporated. For example, transmission probabilities vary over the course of an infection, HIV and other STIs are closely associated in a population, condom use patterns can vary among different dyad groups, and sexual behaviors can exhibit different patterns of mixing. AVERT cannot account for these types of heterogeneity. However, the dynamic character of variables such as patterns of sexual activity, partner change, or condom usage plays less of a role when the modeled period of time is relatively short (1-2 years).

Version 1.0 of AVERT is programmed for use in a DOS environment with an IBM compatible microcomputer. The model is public domain software that may be freely copied. AVERT was developed under FHI's AIDSCAP project, with the support of the United States Agency for International Development (USAID). Copies of the AVERT software and the user's manual have now been widely disseminated to country programs, researchers, and academic institutions.

What's Behind The Model?

The mathematical foundation underlying the AVERT model is a derivation of a probability formula presented by Weinstein and colleagues3. The structure of Weinstein's model is based on the risk of an individual becoming infected through sexual acts with a partner who has been randomly drawn from a population with a given prevalence of HIV infection. AVERT takes this model one step further and multiplies this probability by the number of susceptible individuals in an "at-risk" population. This provides an estimate of new infections. The equation is invoked twice in order to provide estimates for each of the participating sex partner populations. For example, under a given set of conditions, an estimate is calculated for a target population of commercial sex workers and another is calculated for their client partners.

The basic model in AVERT incorporates seven different variables:

• HIV prevalence among sexual partners–p
• average number of sexual partners–m
• average number of sexual acts with a given partner–n
• proportion of sexual encounters in which condoms are used–f
• efficacy of condoms–e
• prevalence of STIs in population–wi
• HIV transmissibility–rgi

The cumulative probability equation is:

To implement the model, a user carries out a first sequence of calculations that yields a probability of risk to the target population (population A) resulting from their having sex with members of the partner population (population B):

The risk to A derives from several behavioral characteristics of A, and from HIV prevalence of and HIV transmissibility from B. The result is multiplied by the number of susceptible individuals in A to yield an estimate of new infections within A. A second set of calculations is then carried out following an identical procedure, but measuring risk to B on the basis of B behavioral characteristics and HIV prevalence of and transmissibility from A.

This result is multiplied by the number of susceptible individuals in B, generating an estimate of new infections within this group.

In summary, once cumulative probabilities are calculated for each study population (and ), those values are multiplied by the corresponding HIV-negative population. These procedures produce estimates of new HIV infections within each group, and the total comprises the estimate for the target and partner populations combined.

AVERT must be run separately for each type of partner population with which members of the target population might be sexually active. For example, if the target population is male factory workers who have sex with both their regular partners and with female sex workers, the model would have to be run separately to get estimates for each group. It is up to the user to select which type of partner populations is relevant to the intervention.

When HIV prevalence and the size of the study population are thought to remain the same between, for example, a pre- and post-intervention scenario, the AVERT program allows the user to compare the two scenarios simultaneously. On the other hand, if the HIV prevalence or the population size vary between the scenarios, then separate runs with the model are necessary to estimate the number of new HIV infections for each scenario. The cumulative HIV incidence rates projected by each run with the model are then the appropriate figures to use when comparing different scenarios.

The following paragraphs provide additional detail on each of the seven variables included in the AVERT model.

HIV prevalence (p)–The HIV prevalence variable incorporates two separate values. When measuring risk of the target population (A), the HIV prevalence of their partner population (B) is used. When measuring risk to the partners (B), the HIV prevalence of the target population (A) is used and calculations are treated independently. HIV prevalence works as a constant variable in all circumstances. This is unlike dynamic simulation models, in which HIV prevalence varies over time within a single scenario. Thus, AVERT should not be used to estimate new HIV cases for time periods of more than one year at a time because new infections are not taken into account in the probability function.

Average number of sexual partners (m)–One value is given for the target population (A) and a second for the partner population (B). These numbers can vary considerably based on the types of populations under scrutiny. The size of the target and partner population is held constant for all calculations when two scenarios are compared simultaneously. While the size of the target population is specified by the user, the size of the partner population is calculated by the program. This is necessary to ensure that the number of sexual acts by both groups are equal.

Average number of sexual acts with each individual (n)–A number is required for the target population (A)–the overall average number of times each target (A) member has sex with each of her/his partners (B)–and for the partner population (B)–the overall average number of times each partner (B) member has sex with each of his/her target population (A) contacts. (It is obvious that the number of sex acts per partner must be identical for any two individuals who have sex with each other.)

Condom use (f)–This variable represents the proportion of sex acts that are protected through use of a condom. Data on different self-reported behaviors might be used to estimate a value for this parameter, depending on what is available from behavioral surveys. For example, if the appropriate data were available, one could base this proportion on the actual number of sex acts with specific types of partners in which a condom is used. In the absence of these data, one might start with the proportion of people in the study population who report using condoms 100 percent of the time with the partner population in question, and then augment that number with a portion of those who report using condoms some of the time. In either instance, this variable is assumed to be spread uniformly across the population.

Condom efficacy (e)–The value assigned to this variable reflects the success rate of condoms in preventing the transmission of HIV during a single sexual exposure. Based on related research, the default value found in AVERT is 95 percent and is used as a constant in all calculations4,5.

Prevalence of STIs (wi)–The STI prevalence variable is separated into four different categories (represented by the subscript i): STIs that cause genital ulcer disease (GUD), non-ulcerative STIs (Non-GUD), a combination of ulcerative and non-ulcerative STIs (GUD + Non-GUD), and no-STIs. STI prevalence estimates are used as a proxy for the proportion of HIV-associated sex acts occurring in the presence of an STI during the modeled period of time. Only one set of values is entered into the AVERT model and applied to the entire study population. If the STI prevalence levels are thought to be different for the target (A) and partner population (B), it is recommended that the higher of the two values be used as the input parameter for the model. This approach allows the model to define reasonably well the number of sex encounters when either partner has an STI. One should keep in mind that the STI cofactor operates in a two-way fashion: through increased infectiousness of HIV-positive individuals and through increased susceptibility for HIV in HIV-negative individuals.

Data entry is restricted to two values: GUD and non-GUD prevalence. These values are multiplied by each other to yield the prevalence for the combination category (GUD + non-GUD), and that value is subtracted from each of the entries to provide the adjusted GUD and non-GUD prevalences. All three of these values are added together and then subtracted from 1 to obtain the no-STI proportion. Each set of calculations–risk to target (A) and risk to partners (B)–is stratified into these four STI groups for the central component of the equation associated with the summation sign.

Results from these calculations are summarized and the final steps in the formula completed.

HIV transmissibility (rgi)–Values for HIV infectivity depend on the specific gender of the study populations. Three basic figures are included as defaults in AVERT: a value for male-to-female, a value for female-to-male, and a value for male-to-male (each of which is represented by subscript g). In a scenario in which the study target population includes female sex workers and their partners are male miners, the initial series of calculations measuring risk to the females uses a male-to-female transmission rate. The second series employs the female-to-male transmission rate to estimate the risk to the miners.

Our transmission rates represent a "best estimate" extracted from published literature6-16, adjusted to the mathematical foundation underlying the AVERT model and restricted to specified types of behavior. For example, the female-male combinations assume vaginal intercourse only, and the male-male combination treats anal intercourse generically, without regard to patterns of receptive or insertive behaviors. These transmission rates also account for the fact that STIs have been shown to enhance the transmissibility of HIV17-26. Ulcerative and non-ulcerative STIs have varying influences, and those effects are incorporated into the model (represented by the subscript i). If one or both types of STIs are identified for the study populations, corresponding transmission rates are invoked for the appropriate stratum calculations (see above). Consequently, each of the basic rates without an associated STI also has three additional variants: one in the presence of ulcerative STIs, one for non-ulcerative STIs, and one for a combination of STIs (this value is the same as that used for ulcerative STIs). A total of 12 transmission rates are therefore available in the program (Table 15-1).

HIV-1 transmission probabilities represent a combination of per contact infectivity and susceptibility, and the probability of HIV transmission is assumed to remain constant over individuals and over time. The selected values represent time-weighted average figures subsuming the different stages of HIV infection. We also assume that exposure to HIV is distributed independently of the presence of a STI. While these choices do not reflect all of the nuances inherent in the HIV transmission dynamics, we decided that a parsimonious model was most appropriate to our purpose.

The model's approach for estimating person to person transmission is based on HIV-1 transmission probabilities per sexual exposure. Unfortunately, only limited data exist on the per-exposure cofactor effect of different types of STIs on HIV transmission8. Most of the published literature on transmission probabilities report cumulative risk estimates, such as risk ratios or odds ratios. Cumulative risk, however, is not easily translated into an increased risk per sex act because it is usually the result of an unknown number of sex acts, of which only a fraction may have occurred in the presence of an STI. Depending on available specific research findings on HIV infectiousness and susceptibility to HIV associated with genital tract infections, in an updated version of the model, additional stratification of the STI cofactor effect estimates may be considered in order to better distinguish between the STI effect on the infectiousness of HIV-infected individuals and the STI effect on the susceptibility of HIV-uninfected individuals in the modeled populations.

Validation Of Avert Estimates

The accuracy of quantitative estimates on the annual incidence of HIV infection generated by AVERT is heavily dependent on the validity of the underlying transmission probabilities per sexual exposure used in the model. AVERT developers, therefore, considered it important to compare the model's estimates with the number of seroconversions observed in a real-life situation. We used the data set of a recently completed randomized controlled trial evaluating the effect of a commonly used spermicide on HIV transmission among female sex workers in two cities in Cameroon between March 1995 and December 1996 to validate the AVERT model27. This study provided detailed data on the number of sexual acts per year with clients and non-clients, corresponding levels of sex acts protected by condoms, and estimates on the prevalence of STIs during the study period (Table 15-2). The HIV seroprevalence among the partners of female sex workers was estimated from 1994 data on male blood donors from the towns Douala, Yaounde, and Ebolowa (HIV positive = 11.5%, n=7,148)28 and from 1996 data on male military collected from 11 army bases (HIV positive = 14.6%, n=1,052)29.

To perform an objective validation of the model, the investigators in the Cameroon trial were asked to provide specific input data without revealing the results of the study. The results of the Cameroon study were deliberately withheld from the authors of this chapter until after they had completed work on the estimates with the AVERT model.

The analysis showed that AVERT estimates of the total number of HIV infections and the annual incidence rate matched quite well with the actual results of the cohort study: 73 estimated new infections compared to 78 actual new infections (Table 15-2). Because the number of infections observed in the study population could not be separated in terms of how many infections resulted from clients, as opposed to non-clients, we did separate runs for the different types of partners with whom the study subjects had sexual intercourse during the year of observation. Interestingly, the analysis suggests that almost 60 percent of the total infections in the cohort of female sex workers were the result of sexual activity with non-clients, most likely a reflection of the different levels of condom use (reported number of unprotected sex acts with non-clients was 2.9 times higher than with clients).

Practical Example From The Field

The following example illustrates the various uses of the AVERT model in estimating the probable number of new HIV infections averted as a result of the actual as well as projected intervention effects of a prevention program among mine workers and women at high risk in a South African mining community.

We have used AVERT to gain a better understanding of the impact of one of the first pilot programs delivering targeted periodic presumptive STI treatment in the developing world. Such treatment has been proposed as an option for reducing STIs in groups at high risk of infection (particularly high-risk women, who often experience no STI symptoms and may not seek treatment otherwise). The program offered free monthly examinations and presumptive STI treatment and counseling, combined with community-based peer education on STI/HIV prevention, to women who trade in sex and others at high risk of STIs in a South African mining community where migrant employees live far away from their families for much of the year. All of the women who used the services were treated with a single-dose antibiotic for the most prevalent STIs in the area. Survey results showed that this approach was effective in reducing STIs, with substantial decreases in STI prevalence among the women using the service and their miner partners after just nine months of intervention.

To analyze this pilot program in South Africa, we constructed scenarios based on reported behaviors and STI test results (Table 15-3). We assumed that the 400 women who used the STI treatment and counseling services regularly had had sexual contact with 4,000 miners living in the nearby hostels. HIV prevalence levels were estimated at 50 percent for the women and 20 percent for the men. Because of the substantial in- and out-migration in the study populations, it was assumed that HIV prevalence rates would remain at this level in the next two years.

After nine months of intervention, investigators estimated that the overall prevalence of ulcerative STIs had dropped from 10 percent to 7 percent and non-ulcerative STI rates had fallen from 25 percent to 17 percent. We observed a 20-percent reduction in the reported number of clients from the miner population, and we estimated that the proportion of sex acts protected by condoms increased from 13 percent to 29 percent. Modeling these scenarios, AVERT estimated that the intervention had averted a total of 237 new HIV infections based on a 1-year time frame: 41 among the women and 196 among the miners.

We also used AVERT to estimate the impact of intervention effects (again calculated for a 1-year time frame) in a scenario in which the following project goals were achieved in the near future (Scenario 3 in Table 15-3): the reduced level of commercial sex acts maintained, 50 percent condom use during commercial sex acts, and an 80 percent reduction in STI infection rates from baseline levels. The results of Scenario 3 presented in Table 15-3 show that the estimated annual cumulative incidence of HIV would decline from 52 percent to 12 percent among the women and from 13 percent to 2 percent among their miner clientele.

We used the potential intervention effects described in Senario 3 as inputs to generate impact estimates of different intervention components. Table 15-4 illustrates that the estimated combined intervention effect (419 averted HIV infections) is not just the sum of the effects of the various single intervention components. The analysis does, however, confirm that targeted presumptive STI treatment was clearly the most effective single intervention component in the selected mining community.

Points For Consideration

The close result in comparing the Cameroon cohort data with the AVERT estimates provides some encouraging validation of the model. How- ever, this result should not be taken as an indication of the model's accuracy in all circumstances.

The structure of the AVERT model reflects several constraints that can influence the validity of the model's estimates. Our intent to keep the model user-friendly and accessible for available project data requires a certain level of conceptual and procedural simplicity. We attempted to achieve this goal while minimizing the loss of model strength. When considering whether to employ AVERT, potential users should keep the following points in mind:

AVERT's static nature–The most prominent limitation in AVERT is its static nature. This model is unable to incorporate change within the time frame considered and does not adjust to the dynamic character of variables and circumstances that are known to exist. Two sets of circumstances can be compared to each other, but only in static terms. This leads to an inability to: 1) account for individuals who became newly infected during the 1-year time frame and who are not removed from the pool of susceptibles, and 2) address the issue of secondary infections that result from those newly infected persons.

AVERT's utility as a program planning tool– Because the selected per sex act transmission probabilities represent time-weighted average figures, our model also may underestimate HIV transmission in settings with an explosive early epidemic and presumably a large proportion of people with a recently acquired infection (and hence increased infectiousness) at that time6,30.

Nonetheless, AVERT can be a useful tool in the design or planning phase of a prevention program because it allows implementing organizations or individuals to estimate the potential impact of a planned intervention program. If project designers have a specific change in HIV incidence rates in mind, the model's outputs can be used to estimate the magnitude of behavior change and improved STI case management that need to be achieved. Conversely, if a particular level of behavior change and STI treatment coverage is envisioned, the designers can see, by incorporating the theoretical numbers into the model, the amount of change in HIV incidence that could potentially result.

In intervention settings with low HIV prevalence, the model may produce relatively small number of HIV infections averted. This may frustrate program managers who have high expectations for the impact of their interventions. A comparison of the estimated HIV incidence rates between the pre- and post-intervention scenarios, however, may provide more insight in the relative success of the prevention program.

AVERT's data requirements: availability and validity– As with any modeling exercise, the biggest challenge involved with AVERT is accurately estimating the input data needed to run the model. The user must find available data from a variety of sources and then derive the most plausible and realistic estimates. Sometimes data are unavailable and sometimes they exist but are unpublished. In other instances, data from different sources may conflict with one another or run counter to conventional wisdom. In addition, the validity of published data may be questionable when methodological constraints exist. However, the data requirements for the AVERT model are relatively minor in comparison with some of the other promoted models2. The main purpose of these more complex simulation models is that they can be used by a multidisciplinary team to conduct complex sensitivity analyses, operations research, and hypothesis testing. In practical terms however, their complexity has greatly limited their use in evaluating specific intervention programs. In contrast, AVERT is similar to Epimodel31 in its ease of use and specificity to the task at hand. With only a brief introduction, a computer-literate user with some background in epidemiology can learn to operate AVERT within an hour or so.

AVERT's data requirements: matching input data to interventions–AVERT is designed to estimate the number of sexually transmitted HIV-1 infections averted due to different types of interventions, assuming that the observed changes are entirely attributable to the modeled interventions. If the project in question was a behavioral intervention, then chances are that behavioral data needed to run the model may have been collected as part of the project. If the intervention was focused on reducing STIs, then users might have good STI data but no behavioral data. In either case, it is incumbent upon the user to find the best data to run the model. Behavioral data may come from surveys conducted with specific target groups, or from behavioral surveillance surveys conducted by the ministry of health and/or other partners. STI and HIV data may come from national or regional surveillance systems, or from isolated studies conducted to gather prevalence and/or incidence data for various risk groups. Although these data may not correspond directly to the populations being modeled, the user must assess the extent to which the data can be used to represent the populations they are interested in. In some instances users may need to adjust available data upward or downward to better represent the populations that users are modeling.

Because the validity of the model's estimates greatly depends on the quality of input data, the accuracy of these inputs will also determine whether AVERT may substantially over- or underestimate program impacts. Given the uncertainty of the input data, we recommend running the model with different likely values. For example, users could include the upper and lower values of the 95 percent confidence intervals from survey data. Such an approach allows users to conduct a sensitivity analysis and they may find that it is useful to present results as a range of model outputs.

Targeted periodic presumptive STI treatment was probably the most effective single intervention component in the studied South African mining community. While this conclusion is relatively valid for the setting considered, one should keep in mind that AVERT does not take into account the decrease in STI rates that might result from condom use and/or partner reduction alone. The model uses STI rates as an input variable for calculating probable HIV transmissions but does not quantify the attributable effects of condom use, partner reduction, and STI treatment on STI transmission and observed changes in STI prevalence.

AVERT's time frame–AVERT estimates are usually based on a 1-year time frame. Because HIV infections through sexual intercourse are the result of accumulated risk exposures, preventing an infection in one year will not ensure that a person remains uninfected in subsequent years. Model estimates of HIV infections averted should be interpreted cautiously, especially in populations with high-risk behaviors where the observed behavior changes suggest that the interventions may only postpone infections rather than prevent them indefinitely.

Conclusion

In interpreting how an intervention will work, models cannot provide simple, exact quantitative predictions with absolute certainty. Models are, by definition, tools for analysis under conditions of uncertainty. They help to identify potentially effective interventions, translate the outcomes of these interventions into impact, and suggest designs for future interventions. With a small number of accessible input variables, AVERT can provide plausible and defendable impact estimates of intervention effects on the reduction of HIV transmission. This straightforward model provides an additional analytical tool for epidemiologists, decision makers, and planners in setting appropriate program priorities. In addition, AVERT may enable program managers to carry out cost-effectiveness analyses of intervention programs that were tailored for specific target populations in various epidemiologic settings.

Acknowledgments

The authors thank Richard Steen for testing the model in South Africa and Leopold Zekeng for making data from Cameroon available for validation.

References

1. Rehle T, Saidel T, Hassig S, et al. AVERT: a user-friendly model to estimate the impact of HIV/sexually transmitted disease prevention interventions on HIV transmission. AIDS 1998;12(Suppl 2):S27-S35.
2. Bernstein RS, Sokal DC, Seitz ST, et al. Simulating the control of a heterosexual HIV epidemic in a severely affected east African city: computational modeling of single vs. combined intervention strategies with the iwgAIDS and simulAIDS models. Interfaces 1998;28:101-126.
3. Weinstein MC, Graham JD, Siegel JE, Fineberg HV. Cost-effectiveness analysis of AIDS prevention programs: concepts, complications, and illustrations. In: Turner CF, Miller HG, Moses LE, editors. Confronting AIDS: sexual behavior and intravenous drug use. Washington (DC): National Academy Press; 1989.
4. Cates W, Stewart FH, Trussell J. Commentary: the quest for women's prophylactic methods–hopes vs. science. Am J Public Health 1992;82(11):1479-1482.
5. Feldblum P, Morrison C, Roddy R, Cates W Jr. The effectiveness of barrier methods of contraception in preventing the spread of HIV. AIDS 1995;9(Suppl A):S85-S93.
6. Mastro TD, deVincenzi I. Probabilities of sexual HIV-1 transmission. AIDS 1996;10(Suppl A):S75-S82.
7. Padian NS, Shiboski SC, Glass SO, Vittinghoff E. Heterosexual transmission of human immunodeficiency virus (HIV) in northern California: results from a ten-year study. Am J Epidemiol 1997;146(4):350-357.
8. Hayes RJ, Schulz KF, Plummer FA. The cofactor effect of genital ulcers on the per-exposure risk of HIV transmission in sub-Saharan Africa. J Trop Med Hyg 1995;(1)98:1-8.
9. Duerr A, Xia Z, Nagachinta T, et al. Probability of male-to-female HIV transmission among married couples in Chiang Mai, Thailand. From: Xth International Conference on AIDS. Yokohama, Japan; 1994. Abstract no.: 105C.
10. Wiley JA, Herschkorn SJ, Padian NS. Heterogeneity in the probability of HIV transmission per sexual contact: the case of male-to-female transmission in penile-vaginal intercourse. Stat Med 1989;8(1):93-102.
11. Satten GA, Mastro TD, Longini IM Jr. Modelling the female-to-male per-act HIV transmission probability in an emerging epidemic in Asia. Stat Med 1994;13(19-20):2097-2106.
12. Mastro TD, Satten GA, Nopkesorn T, et al. Probability of female-to-male transmission of HIV-1 in Thailand. Lancet 1994;343(8891):204-207.
13. Cameron DW, Simonsen JN, D'Costa LJ, et al. Female to male transmission of human immunodeficiency virus type 1: risk factors for seroconversion in men. Lancet 1989;2(8660):403-407.
14. Downs AM, deVincenzi I. Probability of heterosexual transmission of HIV: relationship to the number of unprotected sexual contacts. J Acquir Immune Defic Syndr Hum Retrovirol 1996;11(4):388-395.
15. DeGruttola V, Seage GR III, Mayer KH, Horburgh CR Jr. Infectiousness of HIV between male homosexual partners. J Clin Epidemiol 1989;42(9):849-856.
16. Pinkerton SD, Holtgrave DR, Valdiserri RO. Cost-effectiveness of HIV-prevention skills training for men who have sex with men. AIDS 1997;11(3):347-357.
17. Laga M, Monoka A, Kivuvu M, et al. Non-ulcerative sexually transmitted diseases as risk factors for HIV-1 transmission in women: results from a cohort study. AIDS 1993; 7(1):95-102.
18. Kreiss J, Willerford, Hensel M, et al. Association between cervical inflammation and cervical shedding of human immunodeficiency virus DNA. J Infect Dis 1994; 170(6):1597-1601.
19. Dickerson MC, Johnston J, Delea TE, et al. The causal role for genital ulcer disease as a risk factor for transmission of human immunodeficiency virus. An application of the Bradford Hill Criteria. Sex Transm Dis 1996;23(5):429-440.
20. Plourde PJ, Pepin J, Agoki E, et al. Human immunodeficiency virus type-1 seroconversion in women with genital ulcers. J Infect Dis 1994;170(2):313-317.
21. Kassler WJ, Zenilman JM, Erickson B, et al. Seroconversion in patients attending sexually transmitted disease clinics. AIDS 1994;8(3):351-355.
22. Deschamps M-M, Pape JW, Hafner A, Johnson WD. Heterosexual transmission of HIV in Haiti. Ann Intern Med 1996;125(4):324-330.
23. Mbizvo MR, Machekano R, McFarland W, et al. HIV seroincidence and correlates of seroconversion in a cohort of male factory workers in Harare, Zimbabwe. AIDS 1996;10(8):895-901.
24. Mehendale SM, Rodrigues JJ, Brookmeyer RS, et al. Incidence and predictors of human immunodeficiency virus type 1 seroconversion in patients attending sexually transmitted disease clinics in India. J Infect Dis 1995; 172(6):1486-1491.
25. Celentano DD, Nelson KE, Suprasert S, et al. Risk factors for HIV-1 seroconversion among young men in Thailand. JAMA 1996;275(2):122-127.
26. Plummer FA, Simonsen JN, Cameron DW, et al. Cofactors in male-female sexual transmission of human immunodeficiency virus type 1. J Infect Dis 1991;163(2):223-239.
27. Roddy RE, Zekeng L, Ryan KA, et al. A controlled trial of nonoxynol-9 film to reduce male-to-female transmission of sexually transmitted diseases. N Engl J Med 1998; 339(8):504-510.
28. Tapko JB, Kouinche A, Guiffo L, Kaptue L. Evolution of HIV seroprevalence among blood donors in Cameroon from 1990 to 1994. From: XIth International Conference on AIDS. Vancouver, British Columbia; 1996. Abstract no.: Pub-C-1135.
29. Family Health International/AIDSCAP Project. AIDSCAP project final report 1997. Arlington (VA): Family Health International; 1997.
30. Vernazza PL, Eron JJ, Fiscus SA, Cohen MS. Sexual transmission of HIV: infectiousness and prevention. AIDS 1999;13(2):155-166.
31. Chin J, Lwanga S. Estimation and projection of adult AIDS cases: a simple epidemiological model. Bull World Health Organ 1991;69(4):399-406.