Incidence estimates provide critical insights into the dynamics of the HIV epidemic and are the most direct means of assessing the impact of HIV prevention programmes. The wide-scale implementation of prevention and treatment interventions means that, more than ever, real-time estimates of HIV incidence levels and trends are essential for evaluating the epidemic trajectory, to inform a more efficient and effective response to both the national and global HIV pandemic [1].

Southern Africa remains the region most severely affected by the HIV epidemic. With over six million people living with HIV/AIDS, South Africa has the largest population of HIV- infected individuals in the world, representing a quarter of the estimated HIV infections in sub-Saharan Africa [1]. It is therefore fitting that South Africa has implemented comprehensive national HIV surveillance efforts with the annual antenatal surveys and repeated national HIV household surveys as the core components for monitoring the HIV epidemic. Antenatal surveillance in South Africa has been carried out since 1990 [2] and there have been four large nationally representative household-based surveys, in 2002, 2005, 2008 and 2012 [3].

As the epidemic matured in South Africa and as access to antiretroviral treatment (ART) was rapidly scaled up in the period post-2004, the interpretation of HIV prevalence trends became increasingly complex. ART has increased the survival time of people living with HIV, with the result that HIV prevalence has increased. Hence, measuring ART coverage and estimating HIV incidence at the population level are critical to assessing the impact of treatment and prevention programs on HIV prevalence. Population-based survey methodology has advanced to address these evolving data needs in South Africa. The inclusion of novel laboratory methodologies in the survey protocol has enabled direct estimation of exposure to ART among HIV-positive individuals as well as direct assay-based HIV incidence measures from cross-sectional blood specimens [4]. The prevalence data from the repeated national HIV household surveys were also ideally suited to estimate HIV incidence using a mathematical approach [5].

Incidence assays, which can discriminate those who have been infected recently from others previously infected with HIV, can indicate recent changes in HIV incidence at a population level. Direct assay-based incidence measures using blood samples also provide incidence estimates by risk categories and for selected sub-populations. However, incidence estimation based on testing cross-sectional blood specimens is often established on relatively small numbers of persons that are classified as recently infected, which means that estimates so derived have large confidence intervals. Furthermore, incidence assays should be used in an algorithm where assay-recent specimens are further tested for additional markers such as ART exposure and HIV viral load, to exclude ‘false recent’ results [6, 7].

Mathematical models, such as the Estimates and Projection Package (EPP) / Spectrum [8, 9] and Thembisa [10], can also generate estimates of incidence by leveraging household survey data together with data from antenatal clinic surveillance and information on survival patterns of people living with HIV (and behaviours, in the case of Thembisa). However, there is a risk, in these models, that biases in the estimates can be introduced when certain assumptions do not hold, and as the method is based on using prevalence data, sudden changes in incidence will not be detected rapidly, since such changes only manifest in prevalence levels after a considerable delay. The incidence trajectory in the Spectrum model is fairly flexible, but irregular patterns in the scale-up of ART can induce severe deviations in the estimated incidence trend. Meanwhile, the incidence trajectory in Thembisa is constrained to follow a path dictated by the patterns of risk assumed, for which some reliance is placed upon a simplified scheme of sexual behaviour derived from self-reported data, which may be inaccurate [11]. Finally, simpler modelling techniques based on the synthetic cohort principle, which compare only age-specific prevalence levels at two points in time, can also be used to generate estimates of incidence [12, 13]. However, such methods also have to rely on many assumptions, especially in the era of ART, which leads to highly uncertain estimates and biases where relative levels of incidence across age-groups are not stable.

As each of these methods has different strengths and weaknesses, a synthesis of the different approaches can generate useful insights into the estimates of HIV incidence in a given setting, as well as revealing information about the performance of the methods. In this paper we compare the different methods to estimate HIV incidence in South Africa in order to arrive at an agreement on estimates of incidence. We also compared the performance of the different components of the assay-based incidence testing algorithm to assess the effect of antiretroviral drug testing and viral load testing on direct incidence measures in cross-sectional blood samples.

Prevalence results from the 2005, 2008 and 2012 national HIV household surveys served as key input data to inform indirect model-based incidence estimation. Fig 2 shows the survey estimates of HIV prevalence trends by age group and survey year. There was a significant decline in HIV prevalence among persons aged 15–24 years from 10.3% (95% CI 8.7–12.0) in 2005 to 7.1% (95% CI 6.2–8.1) in 2012 (p<0.001). In contrast, among persons aged 25–49 years, prevalence increased significantly from 20.0% (95% CI 18.2–21.9) in 2005 to 25.2% (95% CI 23.2–27.3) in 2012 (p<0.001).

Given the evolving field of HIV incidence estimation and limitations of current methods for estimating incidence, our results suggest that a synthesis of multiple methods for estimating incidence in the same population is helpful in producing robust estimates of national HIV incidence levels and trends. More confidence can be placed in such results, as opposed to relying on findings from individual methods alone [26]. In the era of rapidly expanding antiretroviral treatment programs, HIV incidence estimation among different age groups is difficult but nevertheless critical for assessing the changing age-specific pattern of HIV prevalence.

Overall, the different incidence estimation methods were in remarkable agreement on 2012 HIV incidence estimates among persons aged 15–49 years. Though the direct assay-based method produced slightly higher estimates of incidence compared with mathematical models and slightly lower estimates of incidence compared to the synthetic cohort, the multi-method comparison shows similar levels and trends in HIV incidence, validating the results of the incidence testing algorithm and highlighting its utility in estimating incidence in cross-sectional settings.

There has been substantial progress over the past decade in the development and evaluation of HIV incidence assays [14, 27]. The LAg-Avidity EIA, used in an algorithm where assay-recent specimens are further tested for HIV RNA levels and for the presence of antiretroviral drugs, is currently the recommended approach to classify recent infections [28]. Our analysis of the performance of the different components of the incidence testing algorithm confirms the utility of testing for ART exposure and viral load to correct for the main sources of false recent misclassifications [29]. We have also demonstrated the relatively poor performance of the LAg-Avidity assay if applied in a single-assay format, requiring a false recent rate correction of 3.1% to reproduce the incidence estimate provided by the full algorithm. Out of 195 LAg-Avidity assay-recent specimens, 106 (54.4%) were re-classified as non-recent infections after testing for antiretroviral treatment exposure and viral load. This algorithm can be performed using dried blot spot (DBS) specimens, which is an important advantage in large population-based surveys [15]. The use of self-report of ART exposure in an incidence testing algorithm may be highly unreliable, as has been shown by several studies [30, 31, 32].

Based on the latest recalibration of the LAg-Avidity EIA in different HIV-1 subtypes, we applied a normalized optical density (ODn) cutoff of 1.5 and the corresponding mean duration of recent infection of 130 days for the incidence estimation [15]. Our assay-based HIV incidence estimates seem to confirm the validity of the selected parameters in the incidence calculation. A recently published assessment by the Consortium for the Evaluation and Performance of HIV incidence Assays (CEPHIA) proposed an alternative MDRI of 177 days and a false recent rate of 1.3% for the LAg-Avidity EIA (1.5 ODn cutoff) in subtype C specimens which excluded treated subjects and elite controllers [33]. However, applying those parameters in the incidence calculation based on our testing algorithm would have resulted in a change of the incidence estimate from 1.72% (95% CI 1.38–2.06) to 0.66% (95% CI 0.04–1.28) in the 15–49 year age group—an estimate that is not supported by any of the epidemiological/mathematical models. A recent revision by the CEPHIA group suggested a MDRI of 140 days and a FRR between 0% and 0.5% for an algorithm that assessed the LAg-Avidity EIA (1.5 ODn cutoff) in combination with a viral load threshold of ≥ 1000 copies/ml [34]. This parameter setup would have produced incidence rates more in agreement with our results, with estimates varying from 1.31% to 1.60%.

We did not include a false recent rate in our incidence calculation, relying on the correction of potential “false recent” results by means of additional testing for antiretroviral drugs and viral load in LAg-Avidity EIA recent specimens. Incidence testing algorithms of this type can reduce the false recent rate to almost zero, as has been demonstrated in samples from individuals in the United States [35]. Although the debate has focused so far on the potential false recent rate, we may also have to consider potential ‘false long-term’ misclassifications associated with this testing algorithm. HIV seroconverters who have been exposed to early treatment may be misclassified as chronically infected individuals on ART, e.g. recently infected pregnant women enrolled in prevention of mother-to-child-transmission (PMTCT) programs or persons on post-exposure prophylaxis (PEP) and pre-exposure prophylaxis (PrEP) regimens [36]. However, the extent of these potential misclassifications in the context of the 2012 national household survey was most likely extremely small. Finally, we should note that antibody-based screening assays are unable to detect acute HIV infections in the pre-seroconversion window. In the recently conducted national household survey in Swaziland, 13 (0.1%) of the 12 338 HIV antibody-negative specimens were identified as HIV RNA positive recent infections [37]. Based on these considerations discussed above we decided not to apply a correction factor to the multi-assay algorithm-derived 2012 HIV incidence estimates at this point in time.

Two mathematical models produced consistent estimates of HIV incidence level and trend, despite important differences in model assumptions and calibration procedures. While the EPP/Spectrum model allows a fair degree of flexibility in the estimation of HIV incidence trends, the Thembisa model estimates of HIV incidence are to some extent constrained by assumptions about trends in sexual behavior. The models also differ in their assumptions regarding ART rollout, with Thembisa assuming greater ART uptake and estimating a greater reduction in AIDS mortality due to ART. While the Thembisa model is calibrated to age-specific HIV prevalence data and recorded death data, the EPP/Spectrum model is calibrated to HIV prevalence data for the entire 15–49 age group. Since HIV prevalence data are not disaggregated by age groups in the model, EPP/Spectrum is not able to capture a scenario in which the incidence trend for 15–24 year olds is different from the trend observed in 25–49 year olds. A future step will be for the models to use age-specific data in their inference of epidemic trends.

All incidence estimation methods produced similar estimates for 15–24 year olds, and were consistent in their estimates of trends. In the population aged 25–49 years, however, the discrepancies between modelled estimates and the 2012 survey-based estimates are more evident. The Thembisa model may have overestimated treatment exposure among older adults aged 25–49 years and, as a result, overestimated the impact of ART on HIV incidence reduction in this age group. In the Thembisa model, 40.6% of HIV-positive adults aged 25–49 were estimated to be on ART in 2012 compared to 31.2% treatment exposure measured in the HIV-positive blood samples [3].

While important differences were observed in HIV incidence levels by sex, age, and time in our analysis, the wide confidence intervals around the incidence estimates suggest that these differences should be interpreted with caution. The synthetic cohort model appears to produce more uncertain estimates than other modelling approaches. This is because the method is unconstrained—i.e. the incidence pattern estimated from the HIV prevalence data does not have to be consistent with any theory or hypothesis about epidemic dynamics. The synthetic cohort approach is more sensitive to sampling errors or other aberrations in the data compared to the mathematical models, which are constrained to follow smooth changes over time or changes that are consistent with an underlying theory of epidemic dynamics. For example, if the previous survey in 2008 slightly underestimated adult HIV prevalence compared to the results obtained by the more representative 2012 survey design, the underestimate would particularly affect the synthetic cohort approach. The incidence estimate among persons aged 15–49 years produced by this method would be 1.5% instead of 1.9% if the bias in the 2008 prevalence estimate was about one percentage point overall and uniform by age and sex. Moreover, many assumptions required by this method are not readily identifiable from the input data, meaning that uncertainties, especially about ART, are propagated to the results.

Our results confirm that South Africa ranks first in the world in the annual number of new HIV infections [1]. The HIV incidence rates presented in this analysis suggest that about 1 000 new infections occurred each day among South Africans aged 15–49 years in 2012. Although the declining trend in HIV incidence between 2005 and 2012 observed among young adults aged 15–24 years is encouraging, the incidence rates still remain at unacceptably high levels, especially among female youth. The current National Strategic Plan on HIV, STIs and TB 2012–2016 states as its primary goal a reduction in new infections of at least 50% [38]. This would require a reduction in HIV incidence well below the 1% level among persons aged 15–49 years by 2016—a considerable challenge given the transmission dynamics that still prevail in the country. The 2012 survey findings indicated that there was a drop in condom use at last sex, an increase in the proportion of people reporting multiple sexual partners, and an increase in the proportion of young women reporting age-disparate relationships [3]. South Africa needs to balance treatment and prevention with a strong focus on the reduction of new HIV infections in the sexually active population.