An absolute rate of HIV incidence

The SDG target for HIV (indicator 3.3.1) is monitored by the number of new HIV infections per 1,000 uninfected population, disaggregated by sex, age, and key populations [10]. A numerical target signifying the ‘End of the AIDS epidemic’ for the HIV incidence rate has been proposed at 1/1,000 person-years (py) in South Africa in the original analysis exploring the powerful effects of antiretroviral treatment in reducing incidence [11]. More recently, the same target value of 1/1,000 py has also been suggested for other high-burden countries [12,13].

However, incidence rates are very different in different populations. A target incidence level for a country with high historical incidence, such as the countries in southern Africa, may not be relevant for a country with historically lower incidence. For example, the incidence in India, taken as a whole, was estimated to be 0.10 (0.08−0.14) / 1,000 py in 2016 [14], and it has never exceeded the 1/1,000 py proposed benchmark [15]. Hence, meaningful targets can only be set in countries and local jurisdictions.

While an incidence rate has been proposed for monitoring the SDG target, both the incidence rate and the number of new HIV infections can be monitored. It is important to realise that, in circumstances of important population growth, absolute numbers of new HIV infections could be increasing while the incidence rate is declining.

An absolute rate of AIDS-related deaths

The level of AIDS-related mortality varies among countries, within countries, by age, and by population groups. AIDS-related mortality in South Africa, for example, peaked in 2006 when 42% of all deaths were attributable to HIV and the AIDS-related mortality rate was approximately 5/1,000 population [14,16,17]. While an AIDS-related mortality of less than 1/1,000 per year may be relevant for the high-prevalence countries in southern Africa, meaningful targets can only be set in countries and local jurisdictions. The proposed indicator for monitoring AIDS-related mortality is the number of AIDS-related deaths per 1,000 population. As above for the incidence metric, both the AIDS-related mortality rate and the number of AIDS-related deaths can be monitored. In the case of important population growth and a decline in mortality, the rate would decline faster than the number of AIDS-related deaths.

A percentage reduction in the number of new HIV infections

For the purpose of the Fast-Track modelling, which aimed to develop programmatic targets and resource needs estimates [18], ‘Ending the AIDS epidemic as a public health threat’ was interpreted as achieving an important reduction in incidence and mortality. Programmatic targets and effect estimates were included in the model for testing, treatment, and viral suppression; condom use; key populations (men who have sex with men, sex workers, people who inject drugs, transgender people, and prisoners) intervention packages; preexposure prophylaxis; voluntary male medical circumcision; and conditional cash transfers. Specifically, the percentage reduction target was set to a 90% reduction in the number of new HIV infections in the entire population between 2010 and 2030. The modelling of the impact resulting from reaching the different programmatic targets achieved a reduction by 2030 of nearly 90% [18]. The 2016 UN high-level meeting adopted values of 90% reduction by 2030 and the interim target of 75% for the year 2020, as well as the corresponding programmatic target values [19].

A percentage reduction in the number of AIDS-related deaths

As for the above percentage reduction in new HIV infections, a 90% reduction in AIDS-related deaths was aimed for in the Fast-Track modelling between 2010 and 2030 [18]. The modelling of the impact resulting from reaching the different programmatic targets achieved a nearly 80% reduction by 2030 [18]. The 2016 UN high-level meeting adopted a target of 90% reduction by 2030 and 75% by 2020 for AIDS-related mortality, together with the corresponding programmatic target values [19].

An IPR

The IPR (incidence:prevalence ratio) is defined as the number of new infections occurring per year in a population divided by the number of persons living with HIV in that same population. When the IPR is below a certain threshold, the epidemic declines—if the IPR is held below that threshold for a long time, the epidemic will eventually be eliminated. The IPR is thus an indicator for the epidemic response ‘proceeding in the right way’, as opposed to it having achieved its goal of ending the AIDS epidemic. It was initially used by researchers in high-income countries [20,4,5], relating it to the effective epidemic number and using it to judge progress in the epidemic and the response. More recently, it has been used to judge progress in the response to the United States epidemic [5–7].

A global benchmark value is obtained by considering the replacement level for the infection: for an epidemic to decline, on average, there should be fewer than one new infection per person living with HIV over the course of their infection. If we assume that newly infected persons survive for D years and there is an equal risk of transmission over this time and that it is the same for all people, then it follows that the benchmark value is equal to 1/D at equilibrium. Making further simplifying assumptions, we can say that D is approximately 33 years (the approximate survival time, accounting for the proportion of PLHIV that benefit from antiretroviral therapy [ART] under UNAIDS’ treatment targets) [9], and so the threshold may be approximately 0.03. We note that this provides only a rough ‘rule of thumb’; a separate paper is in preparation for this collection to explore further the behaviour of the IPR. As the IPR contains elements related to incidence reduction as well as to extending survival, it conveys two major aims of the programmatic response.

An IMR

The IMR (incidence:mortality ratio) is defined as the ratio of the number of people who become HIV infected per year to the number of people among those already infected who die (from any cause) per year. When this number is greater than one, the size of the population of PLHIV grows; when it is less than one, the size of that population shrinks [21,22].

A target of IMR < 1 on its own does not require a reduction of mortality. In fact, dropping below IMR < 1 is possible with a high level of mortality. An additional target therefore needs to be considered in conjunction with the IMR target [21]. A possible target for a direct mortality reduction is that mortality among PLHIV should approach mortality among the HIV-negative population (which will be country and population specific). Indirect targets could relate to ART coverage—for example, ART coverage > 70% [23] or ART coverage > 81% by 2020 and >86% by 2030 [18,24].

There is a close relationship between the IPR (discussed above under the IPR section) and the IMR. At equilibrium, the two metrics become equivalent if the IMR < 1 criterion provides for the mortality rate being below the same level as that implied by the IPR threshold (1/D).

Modelled estimates versus empirical estimates

There is a long history of countries developing modelled estimates of HIV incidence, HIV prevalence, AIDS-related mortality, and all-cause mortality among PLHIV. A major process is undertaken by UNAIDS and partners that involves soliciting expert advice about new model features, software development, capacity building in modelling, and annual country-level development of modelled estimates [25]. This results in the most widely available and consistent set of country-level estimates and time trends in the incidence and prevalence of HIV and HIV-related mortality. In addition, several academic groups are active in this field and publish topic- or country-specific models and estimates [16,26,27] or global and country estimates [28].

Clearly, the quality of modelled estimates depends on the availability and validity of input data. Mathematical models that estimate HIV incidence and AIDS-related mortality may include, besides demographic data and assumptions, the following empirical data: sentinel surveillance prevalence among pregnant women, programmatic data on prevalence among pregnant women, national survey–derived HIV prevalence, numbers of people on ART, national survey–derived incidence, cohort-derived incidence among 15–49 year olds, case reports of new HIV diagnoses, case reports of AIDS-related deaths, and coverage and regimen distribution for prevention of mother-to-child transmission (PMTCT). Data-informed assumptions are typically made about the survival probability of people receiving ART, the effectiveness of antiretroviral regimens to prevent MTCT, and other parameters [29]. Modelled estimates are typically produced at the national level and increasingly for subnational geographic entities: however, currently they typically are not available for specific key population groups.

The major challenge with model-based estimates is that they tend to be most uncertain (least precise) in recent years and are, by necessity, dependent on assumptions for which direct empirical data are lacking. When comparing, for example, incidence and mortality, two quantities are being compared, both of which are being modelled from a dataset that is predominantly informed by trends in HIV prevalence and ART coverage. This can be addressed by using the models to ‘test’ whether there is evidence of a particular target being reached or not, in which the uncertainty attendant in the methods is fully expressed. It may be the case that with current surveillance systems and data availability, one may not be able to make very precise statements regarding meeting particular thresholds as soon as we would like. However, new development in the availability of empirical data—especially measures of incidence and deaths—will bolster this ability.

Empirical data on HIV incidence, AIDS-related mortality, all-cause mortality among PLHIV, and HIV prevalence

Empirical data on HIV incidence are increasingly available but remain rare. Cohort studies that measure the incidence rate directly are costly and time-consuming and may suffer from participation bias [30]. Hence, only a few national cohort studies exist [31,32], with other studies focused on entire populations or subpopulations in specific locations [33,34]. Recently, the cross-sectional population-based HIV impact assessment surveys [35] have been used to generate estimates of HIV incidence (using algorithms that include assays to estimate the number of recent infections) at the national level, although not at the subnational level or among specific key populations. As are cohort studies, these surveys are also subject to participation bias. Separate cross-sectional surveys among key populations are now also recommended to include the measurement of incidence [36], although the body of empirical measurements is currently limited. For both national surveys and surveys among key populations, having a sample size that is large enough to estimate HIV incidence with precision is costly, and this will become even more so as the level of incidence falls to lower values in future years. Nevertheless, the assay-based methods for the measurement of incidence in cross-sectional surveys continue to be investigated and developed [37], and when this information is combined (in models) with other data, the precision of model-based incidence estimates can be improved. Countries with well-developed case-reporting systems can use the annual number of new diagnoses to estimate HIV incidence taking into account other information on the time to diagnosis, underdiagnosis, and failure to report [38,39]. Efforts are ongoing in other countries (including those with the highest levels of prevalence) to develop case-based surveillance and patient-monitoring systems [40]—it is currently unclear whether such systems will allow for the estimation of incidence among key populations.

Comprehensive empirical data on mortality among PLHIV are rare, especially in high-prevalence countries. In countries with well-developed case-reporting systems, data are recorded on all deaths among PLHIV who are diagnosed. In these systems, a proportion of these deaths are registered among people who were not previously known to be living with HIV. Data on all-cause mortality among key populations living with HIV are not available for most countries, although countries with well-developed case-reporting systems that include reporting of the assumed mode of transmission may be able to make estimates based on reported data [39]. In countries with high prevalence of HIV and high mortality among PLHIV, the overall level and the age- and sex-specific patterns of all-cause mortality may be used to constrain modelled estimates [41]. Comprehensive empirical data on AIDS-related mortality specifically are even rarer, especially in countries with a generalised epidemic. Failure to correctly record AIDS-related deaths may be due to stigma [17,42]. In addition, registers may contain the number of PLHIV who have died, but the vital status of individuals lost to follow-up may not be up to date. In countries with well-developed vital registration systems, the failure to correctly record AIDS-related deaths may be due to misdiagnosis of the cause of death [43].

Empirical data on HIV prevalence are widely available. HIV prevalence data are generated for key populations (typically only in specific locations) and for pregnant women in special surveys or through routine reporting systems, especially in countries with a generalised epidemic, and for the entire population from national surveys such as the Demographic and Health surveys [44] and, more recently, the population-based HIV impact assessments [35].