Introduction

More than thirty years into the HIV epidemic, simple, affordable and reliable methods to estimate recent HIV incidence in populations at risk are not available. Recently, this inadequacy has become particularly salient against the background of increasing HIV prevalence observed in young people, especially in young men who have sex with men (YMSM) [1]. For socio-cultural and legal reasons, collection of repeated cross-sectional and prospective serological specimens at younger ages for HIV incidence monitoring is difficult. As a result, such specimens are rarely available, especially of those younger than 21 years of age. Cohort studies provide the best measures of risk and HIV incidence, while taking in account accrual and loss-to-follow-up. In addition, such studies allow causal inference and risk factor assessment for incident HIV infection. However, cohort studies are costly, labor and participant intensive, logistically difficult to implement and sustain, and require clinical infrastructure and an enabling environment. Since incident HIV infection is a relatively rare event, large sample-sizes are needed, while infection is known to occur more often in those not-enrolled or lost-to-follow-up. Nevertheless, cohort studies are essential in guiding and powering HIV intervention studies and in monitoring the overall course of the HIV epidemic. Because of these difficulties, HIV incidence estimates are usually derived from mathematical models [2, 3], or from laboratory methods using increasing and decreasing HIV immune responses to decide on recency of infection [4]. Similarly, epidemic monitoring usually relies on cross-sectional data, such as HIV and AIDS case reports and behavioral and biological surveillance in key populations. While this information is important, uncertainty remains about recency of infection and time-order of events and often data are not age-disaggregated or those of younger ages are under-represented or ineligible. Methods for estimating HIV incidence from cross-sectional data, HIV prevalence surveys or service data have been proposed, including mathematical model-based approaches utilizing a single prevalence survey in near stable conditions [5–7]. Or more generally, incorporating historical data on changes in HIV prevalence and AIDS mortality [8], or from two prevalence surveys in the same population at two different time-points [9, 10]. A related approach is to create an artificial cohort from repeated HIV prevalence surveys along with mortality data to estimate HIV incidence while accounting for population dynamics [11]. If mortality data are not available, the use of the incubation time distribution to AIDS and death may be used as a proxy for this information [11, 12]. Another method is to utilize information from linkable repeat HIV testing and counseling attendees to create an open cohort of persons at risk for HIV infection [13–15]. Recent enhancements to laboratory testing include antibody avidity evaluation and prior HIV testing history, CD4+ cell counts and HIV RNA viral load levels to account for false-recent HIV- positives [4, 16–18]. Some of these alternative methods have been compared for agreement [10, 15]. In daily practice and outside of controlled situations, the predictive value and external validity of available HIV incidence methods have been limited and their use unsuccessful.

During the early 1990s, Osmond and colleagues [19], developed and applied an algorithm for behavioral imputation of person years (PY) of follow-up since first anal intercourse exposure to derive HIV incidence among YMSM prior to enrollment in the San Francisco Young Men’s Health Study (SFYMHS). This method, hereafter called Osmond’s Algorithm (OA), calculated the total number of PY since age at start of anal sexual intercourse and study baseline across all subjects. HIV incidence density was then estimated by dividing the number of HIV-infected YMSM at study entry by this number, times 100. While the behaviorally estimated HIV incidence prior to study entry was equal to its serologically derived equivalent during follow-up (2.6 versus 2.7 per 100 PY), follow-up time was limited (one year) and the number of accumulated PY (n = 192) and new HIV infections small (n = 5) [19]. The non-availability of prospective serological specimens from YMSM is probably one of the main reasons why OA’s reliability has not been further evaluated. Nevertheless, OA was successfully (but of unknown reliability) applied to YMSM in Amsterdam, The Netherlands in 1997 [20], in Bangkok, Thailand from 2003 to 2007 [21], in six cities in China in 2010 [22] and in adjusted fashion to YMSM across the continental United States and Puerto Rico in 2013 [23]. To assess the reliability of OA to estimate recent HIV incidence in YMSM compared to conventional prospective cohort study serology, we applied both methods in parallel against YMSM 18 to 21 and 18 to 24 years old enrolled in the Bangkok men who have sex with men cohort study (BMCS) [24–26]. As part of the BMCS, behavioral and serological data were collected simultaneously in the context of high HIV incidence in YMSM between 2006 and 2014, providing a robust number of existing and new HIV infections to evaluate agreement between the two methods.

Materials and methods

Results

Prior to enrollment, 314 YMSM age 18 to 21 years accumulated 674 PY of follow-up since first anal intercourse exposure. Considering that 50 men had prevalent HIV infection, the behaviorally imputed HIV incidence was 7.4 per 100 PY (95% CI: 5.4, 9.5) (Table 1). Of the remaining 264 HIV-negative YMSM of this age, 54 seroconverted for HIV infection during the study, accumulating 724 PY of follow-up and a serologically estimated HIV incidence of 7.5 per 100 PY (95% CI: 5.6, 9.7) (Table 1). Serologically and behaviorally imputed HIV incidence estimates were similar with an observed rate difference of 0.04 and not statistically significant (95% CI: -2.8, 2.9). Behaviorally imputed HIV incidence did not change after adjustment for self-reported date of HIV-positive testing (n = 3) prior to study entry. Prior to enrollment, 712 YMSM age 18 to 24 years (including men 18 to 21 years analyzed above) accumulated 2143

PY of follow-up since first anal intercourse exposure. Considering that 151 men had prevalent HIV infection, the behaviorally imputed HIV incidence was 7.0 per 100 PY (95% CI: 5.9, 8.2) (Table 1). Of the remaining 561 HIV-negative YMSM of this age, 125 seroconverted for HIV infection during the study, accumulating 1700 PY of follow-up and a serologically estimated HIV incidence of 7.4 per 100 PY (95% CI: 6.1, 8.8) (Table 1). The observed rate difference between the two age groups was 0.31, which was not statistically significant (95% CI: -1.4, 2.0).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Serologically determined and behaviorally imputed HIV incidence in young men who have sex with men 18 to 21 and 18 to 24 years in the Bangkok Men who have Sex with Men Cohort Study, 2006–2014.

https://doi.org/10.1371/journal.pone.0204793.t001

Discussion

In this evaluation of serological versus behaviorally imputed estimations of HIV incidence in YMSM, reliability between the two methods was established. Concordance was found among YMSM 18 to 21 years as well as among those of 18 to 24 years old. This suggests that OA may be applied to a wider age-bracket than previously done [20, 21, 23]. Our study generated robust serologically and behaviorally imputed estimates of HIV incidence. It had eight times the number of HIV sero-conversions compared to the SFYMHS, in which behavioral imputation was originally implemented [19]. Except for ascertainment of age at start of anal intercourse, commonly included as part of cross-sectional surveys among MSM, no additional data are required for behavioral imputation. This method may therefore provide a simple, cheap and reliable technique for monitoring recent HIV incidence in YMSM in settings where longitudinal data are not available.

Cross-sectional HIV prevalence to approximate HIV incidence in young populations has been applied previously to women attending antenatal care [30]. Since early mortality was low in this population, the slope of the prevalence curve could be used to estimate HIV incidence in this group [8]. However, no such institutional data are available for YMSM, who are usually marginalized with limited access to routine primary care services, including HIV testing and counseling. Other methods estimating recent HIV incidence use de-tuned laboratory assays [4, 17] to detect sequentially emerging and declining immune-responses following infection. Yet, these methods have limitations as their HIV infection window is uncertain, such as in the BED-Capture Enzyme Immune Assay (BED-CEIA) [17, 18]. Moreover, the BED-CEIA assumes stability and annualizes HIV incidence, which may not reflect reality. Additional limitations of de-tuned assays are misclassification of late-stage infections as recent, problems with “false-positives”, low-antibody titers in persons on anti-retroviral therapy (ART) and HIV-subtype dependent performance. Attempts to supplement assay-based estimates such as the limiting-antigen avidity assay with information on viral load and ART use to address the false positive problem are not perfect [31, 32]. Mathematical model approaches with one or two rounds of HIV prevalence surveys have also been proposed, but these methods assume the epidemic is in steady-state, irrespective of underlying HIV transmission dynamics [9]. There may also be unknown selection biases in sampling from one round to another, migration, or losses due to HIV-related disease and death. Methods requiring mortality or survival rates are further restricted since these are usually unknown, especially in marginalized populations. In addition, current mortality and survival data are questionable with increased access and uptake of ART over time.

The limited availability, if not absence, of HIV incidence data among MSM in lower- and middle-income countries, has become particularly poignant in the context of the ever ongoing HIV epidemic in this population [1, 33, 34]. This is particularly a problem for YMSM since these persons are at the highest risk but under parental jurisdiction, often do not have access to HIV testing and counseling and are often ineligible to consent for participation in surveillance and research. Moreover, if data are available, they are usually not disaggregated by age. As a result, HIV prevention policy and planning mostly relies on information from adults. This is particularly a problem for YMSM, as age of maturity will not protect them from HIV acquisition, especially when they are young. Our study has several limitations. It was conducted among a self-selected group of YMSM volunteers at risk for HIV infection, hence incidence estimates cannot be generalized to the YMSM population at large. Moreover, a key requirement of behavioral imputation is accurate measurement of when HIV risk begins for YMSM, which may be subject to recall bias and socially desirable answering. Here, start of exposure was measured by age instead of date of first anal sex. While start age is a solid risk-indicator, in many surveys there will be early age outliers, which are unlikely to be representative of HIV risk in YMSM. For these cases a minimum start age may be applied, such as the sample median as was done here and elsewhere [19, 21].

Thirty-five years into the HIV epidemic, simple, affordable and reliable methods to estimate recent HIV incidence in MSM are not available. Recently, this deficiency has become increasingly problematic with the expanding HIV prevalence in YMSM [1]. To stop this from happening, the HIV incidence in the youngest needs to be addressed. In this context, behavioral imputation may serve as a reliable, simple and resource friendly tool to monitor the impact of targeted biomedical and behavioral HIV preventive interventions directed at stemming the HIV epidemic among YMSM in Thailand and elsewhere.

Supporting information

Acknowledgments

The authors acknowledge the support of Ms. Qi Cui and Sandra Kuzmanovska of the Global Fund to fight AIDS, TB and Malaria, Geneva, Switzerland. This work would not have been possible without the contribution of the staff of the Silom Community Clinic of the Thailand Ministry of Public Health–U.S. Centers for Disease Control and Prevention Collaboration. Professor Dennis Osmond (Emeritus) of the UCSF Department of Epidemiology and Biostatistics holds the intellectual ownership of the set of rules evaluated here. This algorithm was named after its creator, for which he kindly gave his permission. This work is dedicated to the life and work of Ms. Supaporn Chaikummao, former manager and head nurse of the Silom Community Clinic, Bangkok, Thailand.