Measuring HIV incidence or the rate of new HIV infections in a population over time is of paramount importance for proper planning and evaluation of HIV prevention programs. Several methods have been proposed for measuring HIV incidence in large areas, yet each presents specific challenges [1], [2].

The original “gold standard” method for measuring population-level HIV incidence is a prospective cohort study that measures the occurrence of new infections in a well-defined HIV-negative population followed over time and tested at regular intervals for HIV infection. These studies, however, are rare, difficult and expensive to implement, and prone to biases that could reduce generalizability of results.

Most developing countries approximate adult HIV incidence using mathematical models that relate observed HIV prevalence to HIV incidence, which make assumptions on the average survival of HIV-infected individuals and the effect of antiretroviral (ARV) treatment on survival [3], [4]. In countries with generalized epidemics, the primary sources of HIV prevalence data for these models are routine unlinked and anonymous HIV sero-surveys among pregnant women attending antenatal clinics (ANC) and nationally representative population-based surveys (NPS) with HIV testing, including demographic health surveys (DHS) and AIDS indicator surveys (AIS) [5]. NPS have also been used to derive age-specific HIV incidence rates in the general population using HIV prevalence data from two sequential surveys in the country [6], [7], [8], [9]. This method has been broadly validated through comparison with cohort measures of incidence and has been applied to several settings where two such surveys exist.

HIV incidence assays are a laboratory-based approach for detecting recently acquired HIV infection in cross-sectional samples of HIV-positive specimens and designed to estimate population-level HIV incidence [10], [11], [12]. These assays are based on the principle that antibody response to HIV infection matures over time and that immunological biomarkers of HIV disease progression can be used to distinguish recent from non-recent HIV infection. The ideal assay has the property that all HIV-infected persons will eventually produce a non-recent test result. The mean time it takes to cross over the defined threshold value defines the assay's duration of recency (ω) [13]. Incidence rates are estimated by combining the number testing as recent on the assay, the mean duration of recency for the assay, and the number at risk for recent HIV infection in an incidence formula [10], [14]. To correct for individuals in a population who fail to progress out of the stage marked as ‘recent’ by the assay [13], [15], the application of statistical adjustments in the incidence formula is required [14], [16], [17], [18]. A critical component to these adjustments is the assay's false-recent rate (FRR), defined as the probability that a chronically infected individual (that is, an HIV-infected individual infected >12 months) will misclassify as recent on the incidence assay. Some HIV-infected individuals that are undergoing treatment with ARVs will also misclassify as recent on the assay as a result of enhanced viral suppression and corresponding decrease in antibody response [19], [20], [21]. Because of the significant impact that ARV use can have on incidence assay test results, all FRR and incidence surveys should have the ability to detect individuals that are currently taking ARVs to appropriately account for these individuals in the analysis. Additionally, auxiliary information on markers of advanced infection that could potentially impact the FRR to a large degree, including duration of HIV infection, CD4 cell counts, and age should also be collected to determine whether FRR varies significantly by these factors [6].

The most commonly used FRR value to date was derived for the BED IgG capture enzyme immunoassay (hereafter referred to as the BED assay) among a cohort of post-partum women followed from 1997-2001 in Zimbabwe. This study calculated a FRR of 5.2% [CI 4.4, 6.1] [16] among 2,749 women with long-term infections that were presumably not taking ARV treatment. The Zimbabwe FRR had been recommended as the default FRR value to use in settings that do not have a locally-relevant estimate of the FRR for the BED assay [18]; however, the universal applicability of the Zimbabwe FRR is unclear given that the FRR was estimated to be 1.7% in neighboring South-Africa [23].

In Uganda, impressive declines in HIV prevalence were documented from peak prevalence in the early 1990s at >10% to an estimated 6% in the early 2000s [24], [25], [26], [27]. More recently, there is evidence that HIV prevalence and incidence are no longer decreasing [28]. In Kenya, HIV prevalence has not changed significantly in recent years, with national surveys showing HIV prevalence at 6.7% in 2003 and 7.1% in 2007 [29], [30]. National adult prevalence was estimated at 7.5% [uncertainty range: 7.0, 7.9] in 2008, using UNAIDS Estimates and Projections Package (EPP) [31].

As countries gather multiple sources of incidence data, there is an opportunity to synthesize these data to determine the best supported level of incidence and incidence trends in populations. We compared several approaches for estimating incidence levels and trends among adults in the general populations of Uganda and Kenya. We used the availability of four different types of incidence methods to draw comparisons between the approaches.

Comparison of assay-derived incidence to modeled estimates of incidence provided evidence that when calibrating assay-derived incidence based on the Zimbabwe FRR of approximately 5%, assay-derived incidence estimates were inconsistent with those obtained by other methods in both Kenya and Uganda. The application of a local FRR of approximately 15% resulted in assay-derived incidence estimates that were reasonably consistent to estimates by other methods in Kenya. In Uganda, assay-derived estimates were two times lower than modeled estimates and similar to cohort-derived incidence reported in the same year. The differences observed were not statistically significant. In the analysis of incidence trends, results obtained by the different methods appeared to correspond fairly well with each other. In Uganda, incidence was stable from 2000–2007. In Kenya, incidence appeared to have declined since 2000 by all approaches.

Comparisons of prevalence and incidence levels in the three NPS confirm that the incidence estimates from all methods (i.e., EPP/Spectrum, the survey-derived method, and assay-derived method using the Uganda FRR) fell within plausible levels for Kenya in 2003 and 2007 (e.g., incidence estimates were 8–15% that of observed HIV prevalence in the same population). In contrast, in Uganda, the two mathematical models of incidence produced plausible levels of incidence (e.g., incidence estimates were approximately10% of prevalence), but assay-derived and cohort-derived incidence estimates were both lower, falling at approximately 4–5% of the prevalence level. The application of the Zimbabwe FRR produced implausible levels of incidence, at levels approximately 30–40% of prevalence, in all three surveys.

These findings confirm that BED assay-based incidence estimates must incorporate a FRR in the incidence calculation to account for false-recent classifications [14], [22], [46], [47]. There is less certainty, however, in choosing which FRR value to apply given the sensitivity of the incidence estimate to the value of the FRR. The Zimbabwe FRR had previously been shown to work well in the setting in which it was estimated in Zimbabwe [16], but did not result in reliable measures of incidence in Kenya and Uganda. Moreover, though the application of a local FRR from Uganda improved the plausibility of the assay-derived incidence estimate for both Kenya and Uganda, the wide confidence intervals around the estimates made it difficult to interpret these findings.

Given the widely differing values for BED FRRs obtained in studies with relatively large sample sizes in South Africa (1.7%), Zimbabwe, (5%), China (6%), and in Uganda (15%) [16], [21], [23], [48], it is clear that more BED FRR studies that are conducted systematically and powered sufficiently are needed to derive this factor in other settings before a determination can be made whether FRR values for the BED assay can be appropriately applied to estimate population-level incidence; whether a local FRR obtained in one country is applicable for all regions within a country and to other countries of close geographic proximity and similar HIV subtypes; and whether the value varies significantly over the course of the epidemic. Another parameter required for estimating assay-derived incidence is the assay's mean duration of recency. Evidence suggests that there may be significant variation in the mean duration of recency across various populations and HIV clades [16], [49]. If used consistently, the value of this parameter should not affect the analysis of incidence trends. However, because the incidence level will be impacted, local estimates of the mean duration of recency may be required to obtain accurate estimates of incidence in a given population. These issues highlight the need for systematic evaluations of the performance characteristics for new and existing incidence assays using standardized methods and well characterized specimen sets. Such an endeavor would require a central specimen repository to be established as a standard resource for these evaluations and to maximize comparability across assays. Specimens in this repository should cover large volumes of specimen panels from HIV seroconverter cohorts and individuals with chronic HIV infection across a wide of geographic settings, viral clades, and epidemic stages [50].

There is clear evidence that specimens from HIV-infected persons that are currently on ARV treatment have a high probability of falsely classifying as recent on an incidence assay and that this error varies significantly by time on ARV [20], [51]. Until an incidence assay that is not impacted by ARV use is available, current incidence assays should only be applied to settings where ARV use can be measured, either on the basis of survey participants' self-report [30] or by using laboratory methods to test for the presence of ARV markers in the blood. Though the latter approach may be more robust than self-report data, limitations still exist that can affect the accuracy of the test such as immediate metabolism of ARVs in the liver. Specimens that test recent on the assay but have confirmed evidence of ARV use can either be excluded from the incidence analysis or reclassified as non-recent on the assay to produce a valid estimate of incidence. While exclusion is an acceptable approach for analysis of assay-derived incidence data, it may result in uncertainty bounds that are wider than necessary [22]. Finally, care must be applied to ensure that the population targeted in the FRR survey and that for the incidence survey are similar with respect to demographics, HIV subtypes, epidemic history, and ARV treatment roll-out; else, the assay-derived incidence estimates will not be reliable. For example, if the FRR is estimated from specimens with persons with longstanding HIV infection and not on ARV, the incidence survey must also exclude such persons from incidence estimation.

The two indirect measurements of incidence in this analysis fell within a plausible range of HIV incidence in both countries. Though the survey-derived model was able to infer incidence using one NPS, this required an assumption of stable HIV prevalence in the preceding 5 years. This assumption was relevant for Uganda given documented evidence of stable HIV prevalence in the general population, but may not be for other countries considering this approach. If stable prevalence cannot be guaranteed for a given setting, it is recommended that this approach not be used until at least two NPS are available [7], [9].

The EPP/Spectrum estimate utilizes routinely collected data from ANC surveillance together with data from NPS to estimate national level adult incidence; therefore this approach remains an attractive method for estimating national incidence in generalized epidemics where these data are likely to exist. The advantage of mathematical models for incidence estimation is that they are easy to use, particularly if the model's input data can be easily accessed and are of good quality. A limitation, however, is that high quality data cannot be guaranteed for some countries due to incomplete reporting and lack of quality control measures in place. Additionally, a degree of uncertainty is associated with the modeled estimates given that they depend both on the structure of the model and on assumptions regarding key parameters which cannot always be determined directly from data for a specific country of interest. Though the assumptions in the EPP/Spectrum model are based on best available data, any errors in the model assumptions (example.g., with respect to survival of HIV-infected persons and ARV use) could impact the quality of the estimates. Further, at the time of writing these models have only been used to estimate incidence by age, sex and location but not by other characteristics (i.e., behaviors, marital status or income level) which may be useful for intervention planning. Finally, because both countries had collected nearly 20 years of ANC surveillance data and had completed one to two national HIV prevalence surveys, the corresponding prevalence and incidence estimates in the EPP/Spectrum models were constrained to narrow bounds which may not reflect the full uncertainty.

Prospective community cohort studies are commonly regarded as the “gold-standard” measure for community-level incidence because incidence can be directly observed in the sample. In this analysis, the main limitation of cohort studies is that they were conducted in limited geographical areas. The Rakai community cohort, for which only early years of incidence were available, reported substantially higher rates of incidence compared to other approaches for estimating population-level incidence. However the reported HIV prevalence level in Rakai in 2002 was nearly two times Uganda's national HIV prevalence in the 2004/2005 AIS. In contrast, the Masaka and Kayunga cohort studies, conducted in areas with lower prevalence than Rakai, reported incidence estimates that were lower than those observed in Rakai but consistent with the measures of incidence obtained with indirect methods for the same time period.

This analysis was subject to methodological issues that may have biased the interpretation of the results. First, the level of the Uganda FRR observed in this analysis was remarkably high. High levels of the FRR will result in large uncertainty in the assay-derived incidence estimate, rendering it difficult to interpret and use these data. Incidence assays that produce consistently low levels of the FRR in a variety of populations are optimal to ensure assays can reproduce valid estimates of incidence for all settings. To guide the development of improved incidence assays, a new target product profile has set the minimum acceptable value of a FRR at <2%, with a coefficient of variation <30%, for multiple HIV subtypes and geographic settings [22], [50]. Second, the Ugandan FRR was derived from adults residing in two geographic regions in Uganda (i.e., rural districts in Eastern and Southwestern Uganda) which may not have been representative of the broader national populations in this analysis and may have impacted the accuracy of the assay-derived estimates. To minimize this bias, the FRR should be estimated in a population that is representative of the one in which the incidence assay will be applied for incidence estimation. For example, if national incidence is desired, the FRR should be estimated in nationally representative samples. Additionally, the FRR may vary by the duration of the epidemic [6], [13], precluding the application of a standard local FRR over time. Though the Uganda FRR did not vary significantly by proxy variables for stage of HIV epidemic [e.g., duration of infection up to 12 years or by age (unpublished data)], given the uncertainty around the FRR, investigators should exhibit caution when applying this value and consider repeatedly measuring the FRR in a representative population over time. If this value is recent or there is evidence that the FRR does not vary over time, it can be incorporated into the incidence formula and expected to result in a significantly improved estimate. Moreover, if improved incidence assays can demonstrate consistently low FRR values in all settings, the need for continued measurement of the FRR prior to conducting incidence surveys will be greatly reduced [50]. Finally, this analysis did not report on age, sex, or geographic estimates of incidence, all of which are expected to vary substantially from national HIV incidence estimates in Kenya and Uganda.

The use of HIV prevalence among young pregnant women aged 15-24 years over time has been used as a surrogate measure for trends in incidence [1], [52]. As the onset of sexual activity in this age group is recent, prevalence is expected to reflect recent infections. However, a limitation in this approach is that it does not inform trends among men nor women aged >25 years. Depending on the surveillance system coverage, the data may not be representative of all regions of the country [53]. Nonetheless, we did find that observed trends in prevalence among ANC attendees aged 15–24 years corresponded well with observed trends in incidence in the overall population obtained through mathematical modeling and published cohort data.

In conclusion, in combination, multiple methods for estimating incidence in Kenya and Uganda appeared to converge in similar trend and levels, yet on an individual basis, each of the approaches have their limitations. It is evident that much work is still needed in the area of assay-derived incidence estimation. Systematic evaluations of incidence assays will help to determine whether this method can accurately and precisely measure incidence. Further, recent infection testing algorithms using a multiple incidence assays in combination with additional clinical (e.g., CD4 cell count, RNA testing), laboratory (e.g., ART testing), and historical information should be explored for improving the accuracy of assay-derived incidence estimates. Pending the development of improved incidence assays, we recommend triangulation of multiple methods for incidence estimation and interpretation of results in conjunction with other epidemiologic data (e.g., HIV prevalence in the same population) to assess plausibility of incidence trends and level in a country and use these data to improve programmatic and policy decisions in the national HIV response.