The analysis consists of two building blocks. First, demographic and HIV-specific estimates and projections were generated using Spectrum software, building on the file underlying the latest estimates on the HIV epidemic published by UNAIDS. Second, projections on the prevalence of NCDs were obtained by extrapolation of age-specific estimates from the GBD 2017 database [24,25]. These two building blocks are then combined to obtain estimates of the prevalence of NCDs among people living with HIV, and of the prevalence of HIV among people with NCDs.

Estimates and projections of the state of the HIV epidemic were generated using the “AIDS Impact Module” of the Spectrum software package [23]. This software is used by UNAIDS and national authorities to derive comprehensive estimates of the history and state of the HIV epidemic from available data, and – conditional on assumptions on treatment access and the path of HIV incidence – can also be used to generate projections. The “AIDS Impact Module” is a compartmental model of the HIV epidemic, in which PLWH progress by age and stage of disease, with age- and sex-dependent transition rates. It is embedded in a demographic framework which pulls relevant data from the UN World Population Profile [26].

Specifically, we build on the estimates by the national authorities and UNAIDS, which form part of the global estimates published by UNAIDS in 2018 [18,27]. The Spectrum work file which had been used to generate these estimates was obtained and is publicly available from UNAIDS (http://apps.unaids.org/spectrum/). Our analysis utilizes Spectrum 5.63, ie, the version which has been used for the UNAIDS 2018 estimates.

The UNAIDS Spectrum work file contains validated estimates through 2017, but the projections (through 2025) contained therein are based on simple extrapolation of most policy variables, and are not (and not meant to be) based on a meaningful forward-looking policy analysis. To adapt the file to our purposes, we extended the time horizon from 2025 to 2050 (by changing the final year in the Spectrum “projection manager” module) [28], and aligned projections on key policy variables with the “National Multisectoral HIV and AIDS Response Strategic Framework 2018 – 2023” (NSF). Specifically, we assume that treatment coverage increases to 95% of adults (ages 15+) living with HIV (91% for men, 98% for women) by 2023, from 84.5% in 2017 (72.4% for men, 93.5% for women), and remains constant from 2023. The NSF does not provide a concrete target for treatment coverage (other than exceeding a benchmark of 90%), but the declines in AIDS-related deaths targeted in the NSF suggest ongoing improvements in treatment access. The assumed increase to 95% in our paper is our interpretation of these policies.

For HIV incidence, we adopted the targets from the NSF through 2023 (declining from 1.1% of adults at ages 15+ in 2017 to 0.2% by 2023, driven by the ongoing expansion of treatment and high rates of viral suppression among those receiving treatment [29]. From 2023, we assumed that HIV incidence declines gradually in proportion with HIV prevalence. The assumptions on HIV were inserted into the Spectrum file using the “direct incidence input” option. Because incidence affects prevalence (and – according to our assumptions – feeds back into incidence), this progress was repeated for several iterations until the projections on prevalence (and thus assumed incidence) converged.

The choice of Spectrum for generating the demographic projections deserves a comment. While there are alternative sources of demographic projections, notably the UN World Population Profile (WPP), Spectrum actually imports its demographic assumptions from a “no AIDS” version of the World Population Profile, while the WPP includes an HIV/AIDS module that is similar to Spectrum. We used Spectrum for the demographic projections to ensure they reflect the latest estimates on the state of the HIV epidemic and assumptions on the policy outlook, and to obtain clear-cut estimates of the impact of HIV.

Age- and sex-specific estimates on the prevalence of NCDs through 2017 were obtained from the GBD 2017 database, a global collaboration to provide estimates of fatal and non-fatal health losses across countries, by sex, age, and – in its latest iteration – 354 diseases and injuries [24,25]. Especially for smaller countries like Botswana, it is by far the most comprehensive source on the prevalence of health conditions. This is achieved by meta regressions over a large body of evidence – from population surveys, medical and insurance databases, and disease-specific academic studies and reviews [25]. As a consequence, GBD estimates – akin to the outcomes of a cross-country regression – may differ from specific observations for that country.

For this reason, we have also reviewed country-specific data which could inform our analysis. The most important alternate data source on NCDs in Botswana is the 2014 STEPS survey which however focuses on the prevalence of NCD risk factors (many of which are self-reported) rather than prevalence of NCDs per se [30]. For prevalence of diabetes, however, this survey reports estimates which are less than one-half of the estimates contained in the GBD 2017 database. We have adopted the estimates of prevalence of diabetes from the 2014 STEPS survey, as the most authoritative nationally validated data (see Appendix S1 in Online Supplementary Document, especially Figure SA1). These national estimates are also consistent with the other international database on the prevalence of diabetes, the IDF Diabetes Atlas [31]. Because diabetes is a major risk factor for kidney disease, we also adjusted estimates on chronic kidney disease for lower prevalence of diabetes (see Appendix S1 in Online Supplementary Document, especially Figure SA2).

We included all “level 2” categories of NCDs covered in the GBD estimates (Table 1); for selected NCDs (especially where there is some evidence on links between long-term survival with HIV and the prevalence of these diseases) we included further sub-categories. There are several approaches which could be used to project prevalence of NCDs. Most results reported in this paper assume that age-specific prevalence remains constant from 2017. An alternative would be to extrapolate prevalence levels based on recent trends in prevalence. We followed the former approach, because age-specific prevalence generally changed slowly, and the differences in projections based on the alternative approaches are therefore small. As age-specific prevalence rates underlying the projections are assumed constant, our projected changes in the prevalence of NCDs can then be clearly attributed to changes in the age composition of the relevant populations. A comparison with findings for an alternative assumption of extrapolating through 2040 based on trends in age-specific prevalence of NCDs in 2010-2017 (Tables S2 and S3 in Online Supplementary Document) shows that the two approaches return similar estimates, and none of the essential findings of the paper would need to be changed if this approach were adopted.

We report estimates and projections on the prevalence of NCDs among PLWH (and among the HIV-negative population), defined as average across sexes and age brackets, weighted by the size of the respective populations. Except for the discussion on direct linkages between HIV status and prevalence of NCDs (see below and Annex), we assume that HIV status and NCD prevalence are uncorrelated (controlling for sex and age), so that the prevalence of NCDs is the same for PLWH and among the HIV-negative population within age/sex brackets. The prevalence of HIV among people affected by NCDs is calculated as average across the population affected by NCDs, weighted by the size of sex and age brackets, again assuming that HIV status and NCD prevalence are uncorrelated (controlling for sex and age).

Where there is reason to expect that age-specific prevalence rates will depart from observed levels or trends, it would make sense to accommodate such changes in the projections. For example, policies towards reaching the 25×25 targets (of achieving a 25% reduction in premature mortality from NCDs by 2025) might result in such changes [32]. We have not attempted such projections. The 25×25 framework involves a mix of measures to reduce incidence of NCDS and to improve survival with NCDs, and cannot easily be mapped into assumptions on prevalence, and modelling efforts inspired by 25×25 targets focus on mortality rather than prevalence of NCDs [33]. Similarly, the national NCD strategy describes strategies to mitigate NCD risks and to improve health outcomes among people affected by NCDs, but includes a “no increase” target for prevalence of diabetes and obesity [20].

We do however address the potential implications of direct linkages between HIV and NCDs. If HIV or exposure to ART results in an increased risk of developing some NCDs, then this NCD will be a more common co-morbidity among PLWH, and HIV will be a more common co-morbidity among people affected by this NCD than among people not affected by it. We therefore provide alternate estimates, introducing a multiplier that differentiates prevalence of selected NCDs among PLWH and people not living with HIV, controlling for sex and age group, for ischemic heart disease and diabetes mellitus where empirical evidence is available, including studies from outside high-income settings (see Appendix S3 in Online Supplementary Document) [34-42]. For the projections, we do not distinguish PLWH on ART and without treatment, in light of the very high treatment coverage rate achieved and targeted in Botswana, especially for PLWH reaching old age. For this reason, we also do not account for the effect of HIV or ART on the body mass index (a risk factor for some NCDs). As treatment results in recovery of weight losses [43], and most PLWH in our projections will have been on treatment for a number of years (especially those at older ages when NCDs become more common), differences in body mass index (BMI) between HIV-positive and HIV-negative people will arguably not be a significant factor in the evolving burden of NCDs in Botswana.

We did not include a coefficient to account for the interactions of HIV and cancers in the calculations because – unlike for AIDS-defining cancers – the empirical evidence is diffuse and not clear-cut with regard PLWH who are receiving ART [44,45].

Some of our results mirror findings from the literature on the “greying of AIDS” and on the increasing role of NCDs among PLWH. As the population living with HIV ages, the prevalence of those NCDs which are more common at older ages increases among PLWH, and the health needs of PLWH become more complex in the presence of these co-morbidities [5-7,11].

The comprehensive demographic and population-level approach adopted here – focusing not only on PLWH but covering both HIV-positive and -negative people, generates important new insights for policy and practice. From this population perspective, the increase in the burden on health systems associated with the aging of the population living with HIV is not clear-cut – the health needs of PLWH become more complex, but HIV prevalence declines.

An additional factor that is often overlooked is past HIV-related mortality and its demographic implications. Cohorts reaching old age now have been decimated by AIDS-related mortality in middle age as many members of these cohorts contracted HIV and developed AIDS before ART was widely available. For this reason, HIV has been and is slowing down population aging, and it has reduced the share of the population at old age and the prevalence of age-related NCDs. However, as cohorts who have benefitted from the comprehensive treatment scale-up early on reach old age, this effect will dissipate, and the growth of the population at old age and of age-related NCDs will accelerate, – in Botswana, most notably between 2030 and 2040 (Figure 1, panel A).

In addition to the growing health needs of the population living with HIV, our findings introduce a second dimension through which HIV contributes to emerging health systems challenges. As observed elsewhere, population aging – irrespective of the impacts of HIV – will result in increasing prevalence of most NCDs in Botswana. We demonstrate that, even though HIV prevalence declines overall, the prevalence of HIV among people affected by many of the most important NCDs will increase or barely change. Because of the changing age profile of PLWH, HIV will thus be concentrated and persist in the growing populations facing the most complex health needs.

Some of our methods and results reflect the fact that estimated and projected treatment coverage in Botswana is already high, especially in older age cohorts where NCDs become more common. In countries where the scaling-up of treatment occurs later or is less complete, the projected acceleration in the growth of the older population would be delayed or less pronounced, and transitional factors associated with the scaling-up of treatment (recovery in BMI, reduced prevalence of AIDS-defining cancers) might have to be considered. However, the situation in Botswana resembles that in many other Southern African countries, in terms of very high HIV prevalence, a relatively high life expectancy before the onset of the HIV epidemic, and high ART coverage. More generally, a confluence of a high presence of HIV and an increasing burden of NCDs, coupled with a health system that is already stretched, is observed in many African countries, suggesting that our findings are relevant more widely [13].

Apart from the question as to what extent the results generalize, the principal limitation of the analysis arises from the simple assumptions on age-specific prevalence of NCDs and on HIV-NCD linkages. In this regard, it is encouraging that the results are robust to the two alternative assumptions on the trajectory of age-specific prevalence of NCDs. One item that will benefit from further work (but has not been captured for lack of solid evidence) is the interaction between HIV care and treatment and the onset of NCDs. To the extent that choices in HIV care and treatment affect the propensity to develop NCDs, our analysis may not capture fully the policy space for managing and shaping the joint HIV-NCD burden.

In summary, health systems challenges posed by HIV/AIDS and its intersections with NCDs will grow and persist. The demographic impacts of HIV slow down overall population aging and thus the growing burden of many NCDs at present, but the lagged impacts of large-scale HIV treatment will accelerate population aging and increase chronic disease burden from about 2030. While HIV prevalence declines, the needs of PLWH become more complex. And HIV prevalence among the growing populations affected by various important NCDs will not decline for decades, leading to the persistent health systems challenge of how to successfully treat and care for the large numbers of patients suffering from both HIV and NCDs.