In the mid-1960s while at Pembroke College (Oxford) I used to spend time, rather too infrequently I confess, in the college library. It was a small space, made smaller by the fact that the bottom floor could not be used, for the good and sufficient reason that it was so full of ancient tomes that there was no room for people. And there was no way, either, of accessing any book further than one foot from the door. Some years later a wealthy American made a large bequest to Pembroke and a huge new library was built and suddenly there was room not only to store the old material but also for scholars to research material going back many centuries.

The problem of space to store, and also have accessible, hard-copy data has always been a problem. But as long as the data are stored there can come a time when they can be retrieved and used. Of course things don’t always turn out that way …. In the 1970s, while working on tsetse flies in Rhodesia, we had occasion to move offices and questions arose about what to do with the mountains of old field records of fly catches, and laboratory records of blood meal analyses of what the flies had been feeding on. There was no sensible place to store these records and I confess to being party to the young and foolish decision that the methods used to collect the data were sufficiently spurious that the data were of dubious value: and we burned the lot. And I then almost immediately realised that I really needed those data. It was a harsh lesson, which has influenced my attitude to data ever since.

Of course starting in the 1980s it seemed that all of these problems had been solved. Suddenly we had computers with tape storage facilities, and then floppy disks, then “stiffies” (dread word) and CDs and then the web with gigabytes, terabytes and now, I see from Google, even yottabytes of storage space. Almost infinite amounts of data, it seemed, could be stored safely, for ever, and at no cost. Aaah, yes. Except that the tapes full of data that I brought back from California in the early 80s were unreadable only five years later, and the second generation of desk-top computers could no longer read floppies, and the next couldn’t read stiffies, and even before increasing numbers of brave new storage devices became obsolete they, unaccountably, became corrupted. But never mind, onward and upward, we had hard disks that could hold 20K of data – imagine! And government departments with too much paper to store, copied the data onto their computers – and then followed my example and burned the paper.

Fires can of course happen by accident, but “book burning” is generally a conscious decision. To destroy all of the valuable data in the old Pembroke Library somebody would have needed to physically move all the books out of the basement, make the pile and light the mach. The destruction of data on computers, however, can happen much more insidiously. The person who collected the data moves on; the next person who takes the job has no idea what this, essentially “invisible”, material is all about and either deletes it all at a single key-stroke or, more likely, doesn’t think to make copies when the computer becomes obsolete and is tossed out.

Even the most modern web sites, which are in some senses independent of the individual hardware on which we collect the data, need to be maintained. Somebody has to take care of the site, somebody must pay, somebody must be paid, or the material simply vanishes – with no smoke and no flame to warn us of the crime that has just been committed. Perhaps in a matter of days, or months, or years, or perhaps, if we are lucky, a few decades later the material will be gone for ever. It seems we have no “Pembroke basement” where we can leave modern information, safe and undisturbed for centuries.

These are not simply theoretical worries. Certain death records in our own country were “computerised” some years back and the hard copies were about to be destroyed when a vigilant hoarder asked if she might take them. The custodians of the material perhaps thought her a little quaint but agreed to her request. Not many months later they came back to her, cap in hand, to ask if they might consult the hard copies. It turned out that somebody had hit the wrong key on a computer and all of the carefully entered electronic data had gone up in metaphorical smoke. In at least one neighbouring country the outcome was not so happy; when the computer records went that was the end and there is now no record of death rates in anything other than the immediate past.

This is one facet of the problem of information retention. There is another, and even more worrying, side to the problem and this concerns the collection of information which is simply never used. Such information might as well have been burned or, in fact, never collected since data collection is generally an even more expensive undertaking than data storage.

And this problem brings us back to the matter of ART and HIV incidence. The essential idea of Treatment as Prevention (TasP) is that, in settings where we have a high HIV prevalence, treatment should start immediately for those people who are HIV positive or, in a variant of the idea, should be offered as a prophylactic to those at highest perceived risk even before they become HIV positive.

The idea of all of this is to decrease HIV incidence. The question is how should we measure this incidence; how should we decide whether the increased use of TasP is indeed resulting in decreased HIV incidence? To this end there have been strenuous efforts made in recent years to come up with fancy ways of estimating HIV incidence from the more easily estimated prevalence levels, or death rates, or using bio-markers. In almost every case when workers have presented such approaches there has been a nod to the fact that there is a “gold standard” of incidence estimation which involves the follow-up of cohorts of individuals. Whereas this approach carries with it certain biases and assumptions, the major reasons given for not using the approach more frequently are generally cost and logistical difficulties.

But the common thread in the two variants of the TasP approach is that they will require vastly increased levels, and increased regularity, of HIV testing. This is already happening to a huge extent in countries such as Botswana, where persons attending clinics are tested on an opt-out basis. In other words, as part of the TasP process, it will be necessary to gather the very follow-up HIV testing history required for incidence estimation using the gold standard approach. Unfortunately nobody (yet) seems to be collating and using these data to see what incidence estimates emerge.

There will of course be problems because the samples of people are not randomly selected. On the other hand, the whole target of TasP, as applied in high prevalence settings, is that there should be an effort to test as large a proportion of the sexually active population as possible. The more closely this ideal is approached the better should be the resulting HIV incidence estimates. For example, if time-series information is available on the HIV status of every pregnant or postpartum woman every time she attends a pre- or post-natal clinic it would at least be possible to estimate accurately the incidence among pregnant and postpartum women at close to a population level.

What stands in the way of making sensible use of such data is the current obsession with secrecy when it comes to the collection of data on HIV, unlike the results on any other disease. Nonetheless it should be possible to devise ways whereby, without compromising the patient, serial HIV status records from individuals can be linked together and used to estimate HIV incidence.

If we can’t link the records in this way then we might as well burn them, or bury them in the basement of the Pembroke library.

BED testing for HIV incidence estimation

Brookmeyer and Quinn (2) were the first to propose the use of cross-sectional surveys to estimate HIV incidence, using a biomarker-based approach. Their method consisted of observing a biological marker indicating an immune system response to early infection, to classify individuals as either recently infected or non-recently infected and to use the biomarker results to estimate incidence.

Prior to the usage of a biomarker to estimate incidence, it is necessary to estimate the mean duration of the time individuals spend in the recently-infected state (“window period”). Unfortunately, the mean window period used in the publication quoted above is very short. This implies that unreasonably large sample sizes for the incidence cross-sectional surveys are needed to obtain accurate estimates. In order to improve the properties of the biomarkers, Janssen et al. (3) proposed a method based on the use of ‘detuned’ assays to detect recently infected individuals, which facilitates better precision in the incidence estimates because of a longer window period. The use of assays did not prove to be reliable because of subtype diversity which causes variability in immune response. The BED assay, which is a capture enzyme immunoassay (CEIA) based on protein sequences from the B, E and D HIV subtypes (4), was developed to improve biomarkers characteristics. However, in 2005, the UNAIDS Reference Group on Estimates, Modelling and Projections warned against the usage of BED to estimate HIV incidence rates. They called for the development of additional laboratory and modelling methodologies (5). However, currently the BED is the most widely used incidence assay and a systematic review of the BED incidence assay reported that it could produce accurate estimates of HIV incidence rates, if correct parameters were used (6).

Two of the current challenges in using HIV incidence assays to characterize HIV incidence rates are the knowledge of the BED window period and the misclassifications. The number of HIV-infected persons falsely identified as recent seroconverters, which is the main source of misclassification, depends on the proportion of HIV-positive participants whose infection duration exceeds the BED window period. This number, in absolute value, is lower among young people of a sub-Saharan African setting with high incidence rates, in comparison with populations with wider age ranges. In fact, in such settings, young people have been exposed to HIV for a short duration because of their recent onset of sexual activity and amongst them the fraction of HIV-positive individuals on antiretroviral drugs or having low CD4 count is lower. Thus, the BED incidence assay is optimised when it is used among young people of sub-Saharan African settings with high incidence rates.

Moreover, although the conventional cut-off value for the BED assay is 0.80, corresponding to a BED window period (W) of about six months, Fiamma et al. showed in an empirical study that higher cut-off values of up to 1.89, corresponding to a W of about 15 months, can be used among young people in South Africa (7).

The study further suggested that the BED incidence testing may be used to assess the effect of an intervention. More precisely, the authors demonstrated that the protective effect of male circumcision could have been calculated using only blood samples collected from participants at the last follow-up visit of the Orange Farm male circumcision trial (7).

BED testing for HIV incidence rate ratios estimation

The generalisation of this result was investigated in a recent study (8). The authors of that study examined the capacity of the BED incidence testing to estimate the effect of a prevention intervention and provided a tool to calculate statistical power, when used among young people of sub-Saharan Africa. They designed a study where a fraction of a population was offered an intervention aimed at reducing the effect of HIV acquisition. Then theoretical calculations were performed, which led to a formula giving the maximum likelihood estimator of the effect. Their formula provided an unbiased estimate of the effect when the population size was large. The theoretical power which they called “BED theoretical power” could then be derived by using either statistical tools or simulations. This was compared to the cohort power, which is the statistical power to detect a significant effect in the case of a classical cohort study with the same duration as the window period.

Because the BED theoretical power estimate obtained from their theoretical calculations is not usable in practice, the authors proposed a method to estimate HIV IRR from empirical data. They simulated random samples of individuals belonging to the intervention and control groups. Each individual was characterised by their time since sexual debut, HIV status and the variable indicating recent seroconversion status (Yes or No for those HIV-positive and NA – not applicable – for the others). Data from this simulated population were analysed using a Poisson-log-linear model to estimate the value of the intervention effect. The process was repeated 10 000 times and the “BED practical power” estimated using these simulations.

Numerical simulations were performed using values from published data. The baseline values for HIV incidence rates amongst men and women were 2.1% and 5%, respectively. The intervention was assumed to reduce the HIV incidence by 60% which corresponds to an effect of 0.4. In almost all the simulations, the size of each group was 1500 and the (maximum) duration of sexual activity was 6 years. The characteristics of the BED assay were assumed to be perfectly known in the simulations. These simulations showed that when the short-term specificity is equal to the sensitivity and equal to 0.87, and if the long-term specificity is equal to 0.96, then the BED theoretical power is 0.62 for men and 0.91 for women, which is 94% and 97% of the cohort power for men and women, respectively. They also showed that: a) the power increases with increasing HIV incidence rate, increasing BED window period, increasing sample size or increasing duration of the intervention and b) the power decreases with increasing time since sexual debut or decreasing intervention effect, decreasing specificitiy or sensitivity of the assay.

The practical estimations of the effect, which consisted of estimates obtained using a Poisson log-linear model, were relatively close to the true value. These estimates were the poorest when the characteristics of the assay were not known with great precision or when the Poisson model was not weighted. Results from these simulations also showed that the BED practical power for men and women are both roughly 75% of the BED theoretical power for men.

Finally, the method was applied to empirical data collected at the last follow-up visit of the Orange Farm male circumcision trial (9). The estimated effect was in strong agreement with the values obtained from classical statistical analysis. The effect was underestimated when not corrected for the long-term specificity and misclassification.

Overall, the study demonstrated the ability of the BED testing to a) reliably measure the effect of an HIV intervention aiming at reducing the HIV incidence rate, with classical analysis of individual data, correcting for misclassification and b) lead to a statistical power close to the power obtained in classical cohort studies conducted among samples of the same size as the cross-sectional survey. This suggests the need to change the assumption according to which HIV incidence assays are not of practical use because they require a very large sample size.

• To determine if ART in HIV-discordant couples is associated with lower risk of HIV transmission to an uninfected partner compared to untreated discordant couples.
• To assess if HIV transmission is lower in couples in which an infected partner has CD4 cells ≥ 350 cells/µL.

Lower risk of HIV transmission in treated couples

One randomised controlled trial (RCT) with 1750 HIV discordant couples (7) and seven observational cohort studies with 9791 couples were included in this review. The RCT was conducted in nine countries (Botswana, Brazil, India, Malawi, Kenya, South Africa, Thailand, United States of America and Zimbabwe). While this trial mostly included heterosexual partners, homosexual partners were also included. Most of the cohort studies were conducted in African countries and focussed on heterosexual partners.

The included studies identified 464 episodes of HIV transmission, 72 among treated couples and 392 among untreated couples. In the RCT, the risk of HIV transmission was reduced by 96%. All HIV infected partners in this study had CD4 cell counts of 350-550 cells/μL at baseline.

In the observational studies, the risk of HIV transmission was 66% lower in treated compared to untreated couples. There were major differences between these studies, as two had inadequate data corresponding to the time patients were followed up. After excluding these two from the analysis, the risk of transmission was 84% lower in untreated couples. Among couples in which the infected partner had a CD4 count greater than 350 CD4 cells/μL, the risk of transmission was 98% lower among treated couples. In this subgroup, there were 61 transmissions in untreated couples and none in treated couples. The safety data is extracted from the HPTN052 study which was stopped early. Fourteen per cent of participants who received ART, regardless of their CD4 count, had one or more severe or life-threatening events, which may or may not have been causally related to the antiretrovirals, suggesting no increased risk associated with starting ART at high CD4 count. In contrast, there was a relative increase in grade 3 or 4 laboratory abnormalities among participants receiving therapy with CD4 counts greater than 350 cells/μL (27%) when compared to participants receiving treatment at lower CD4 counts (18%).The study had been powered to evaluate HIV incidence, and was stopped before all safety data was collected, therefore we cannot draw conclusions from the safety results provided.

Implications for practice

Based on the evidence provided by one randomised controlled trial and seven observational cohort studies, ART has been shown to be a potent intervention for prevention of HIV in discordant couples. An important question from a clinical standpoint is whether being in a serodiscordant relationship and having a CD4 count greater than 350cells/μL should be an additional indication for ART under WHO guidelines. European and U.S. guidelines already allow for starting at a CD4 count of up to 500cells/μL routinely and even higher for certain subgroups and based on clinician judgment. The included RCT provides definitive data demonstrating a large positive benefit. Therefore, patients beginning ART may also be informed that adherence to ART can also reduce their risk of transmitting HIV to their uninfected partners. A related policy question is how much effort should be focused on treating individuals with CD4 counts greater than 350cells/μL when access to ART for persons with less than 350 CD4 cells/μL is far from universal. Significant questions remain about the durability of protection, cumulative antiretroviral toxicity, when to start treating an infected partner (for instance, at diagnosis or at a specific CD4 or plasma viral load level) and transmission of ART-resistant strains to partners. The success of this intervention likely relies on good adherence, especially in stable couples. Programmes should be designed that include counselling, support, follow up and mutual disclosure, as these components may have a role in supporting adherence. In addition to ART provision, limitations in resources needed to implement such expanded ART indications must be addressed.

Implications for research Additional data are needed on durability of protection for uninfected partners, adverse events associated with initiation of ART on individuals with CD4 counts greater than 350cells/μL, including effects of longer-term ART, the potential for earlier development of antiretroviral resistance (resulting in a need to change regimens prematurely) and HIV morbidity, quality of life and the potential for risk compensation. There are multiple opportunities to examine these issues in existing cohorts.

Scientists at SACEMA have been in the forefront of those arguing that the time to end AIDS is now (2) and the way to do this is through the strategic use of potent anti-retroviral therapy (ART). The road ahead will be long and hard and much still needs to be done. If we are to increase the number of people in the world who are on ART from the present 5 million to 15 million by, say, 2015 and to 30 million by 2020, many operational challenges will have to be understood and met. Here we outline the most important issues that need to be explored if treatment-as-prevention is to become a reality and if we are to end AIDS.

Drug supply

Drugs will have to be manufactured in sufficient quantity. This is a matter for negotiation between the funding agencies and the drug companies and the possibility of a large, guaranteed market should make it possible to reduce drug prices even further. Careful thought needs to be given to the optimal first and second line regimens in developing countries with special attention being paid to the needs of pregnant women and children.

Drug delivery

Delivering 15 to 30 million daily doses of drugs will be challenging, but the drugs appear to be chemically stable and it will be important to ensure that adequate supply lines are in place. This will be a matter for negotiation between the funding agencies and national governments.

Testing

People need to be tested for HIV on average once every one to two years. There are many ways in which this can be done using provider initiated counselling and testing; snowball sampling; outreach/support groups; community workers; voluntary counselling and testing in mobile clinics; home based testing; couples counselling; methadone programmes; needle exchange programmes; and campaigns. How it is done will vary from place to place but given the will and the imagination this should not be an obstacle to rapidly scaling up ART. Success will depend on the extent to which local community members are actively engaged in supporting testing.

From testing to treatment

At present there is considerable attrition between testing positive and starting treatment and in some settings as many as 90% of those that test positive never start treatment. This happens mainly because of the complexity and difficulty of carrying out CD4+ cell counts, the reluctance in some places to use rapid tests, and the need for poor people to make repeated visits to distant clinics only to be told that their CD4+ cell count is still too high. By abandoning CD4+ cell counts, except for research and in so far as they may be used to monitor progress, the decision process for starting treatment can be simplified, streamlined, and cheaper, and the attrition can be avoided.

Compliance

Once started on ART it is essential that people comply with their drug regimens (3); we need to reach levels of at least 80% to 90%. While many studies show that compliance in Africa is generally better than in developed countries, mainly because HIV positive people in Africa are often poor while people in developed countries may be intra-venous drug users, homeless, alcoholics or otherwise marginalized people, ways must be found to improve compliance above currently reported levels. This will depend largely on community awareness of the importance of compliance and on community support for those on ART.

Viral load suppression and viral rebound

Good compliance is needed for the benefit of the individual person taking ART, to minimize residual transmission and to reduce the development of drug-resistance. For this reason viral load monitoring, perhaps of a sub-set of patients, especially during the first year of ART, will be needed, and support should be given to scientists developing cheap, point-of-care viral load tests. In a drug trial (4) carried out in the USA the median viral load fell by 100 times after one month, 10 thousand times after one year and 100 thousand times after seven years to a level of between 2 and 20 virions per millilitre of plasma. Good viral load suppression can be achieved, but we need to find ways to achieve it in often marginalized communities.

Drug resistance

While drug resistance was of great concern in the early days of triple therapy (5-7) the worst fears have not materialized (8) and there is strong evidence that with good coverage and compliance, and using appropriate drugs, drug resistance will decline rather than increase (9).

Population level impact

While there is evidence from observational studies that the increased availability of ART has led to population level declines in the incidence and prevalence of HIV (10-14) this needs to be confirmed and measured using both population level measures of incidence and measures of transmission among discordant couples. While a number of studies are being planned to explore this in detail (15), it should become part of routine monitoring of scale-up programmes.

Stigma and discrimination

Stigma and discrimination threaten the success of HIV-control programmes and these issues must be confronted. However, the widespread and ready availability of ART should make it possible to ‘normalize’ HIV infection especially if it is clearly understood that those on ART will not infect others, if they are fully compliant, and that they may be expected to live a full and productive life. Dealing with stigma and discrimination will depend largely on the extent to which community members are engaged and the programme is used in a positive way to create jobs and stimulate local economies. Fortunately, there is evidence that by making ART more widely available leads to a reduction in stigma (16).

Seventeen percent of all those living with HIV live in South Africa. The burden that this has placed on our still nascent democracy is almost beyond words. But we can now end the AIDS epidemic if we have the will, the commitment and the imagination to do it.

The rationale by some health practitioners is that the risk of acquiring HIV infection is low, in the region of 0.04% and 1.7% per sex act (1), and that an increase of this low level of risk from, say 1 in 1000 to 2 in 1000, makes little difference to the infection risk of an individual for a given sex act, and by extension, does not appreciably increase incidence at a population level.

Health authorities often invest in barrier method preventative measures because the infection risk per sex act is reduced by orders of magnitude (2). However, the case for investing in factors that may reduce HIV infection risk by a mere 2-3 fold is less clear. In the case of female hormonal contraception, for example, should the South African government alter its policy on administering contraception if such contraception is shown to double the per-sex-act risk of infection?

The question of whether or not doubling or tripling infection risk warrants concern is best considered with a straightforward mathematical description of infection probability. Let R be the overall risk of infection, let P be the probability of HIV infection per sex act, and let n be the number of sexual events per day. With these parameters, the risk of infection for a seronegative individual performing one or more sex acts with an HIV positive partner, is given as follows:

Figure 1 shows a plot of the resulting probability of infection for three different risk scenarios. Line 1 (red) shows baseline risk (assumed to be 0.0016 probability of infection per sex act), while line 2 (blue) and line 3 (green) show a 2 fold and 3 fold increase on the baseline, respectively. In each case the probability of infection begins to approach unity as the number of sexual events increases.

Figure 1. Probability of HIV infection versus number of sex acts. Infection probability per act is assumed to be 0.0016.

This simple probabilistic treatment of sexual events shows that a 2 and 3 fold increase in infection risk does indeed translate into an appreciable increase in infection likelihood. The risk is perhaps best contextualised by making assumptions about the number of sex acts per day, and then examining the subsequent infection risk over time. When assuming one sex act every five days, for example, an individual in the baseline risk group has an 11% chance of contracting HIV over the course of a year and a 95% probability of infection within 5 years of sexual activity. Individuals at double and triple the baseline risk have, respectively, a 21% and 31% probability of contracting HIV within the year, and a 95% probability of becoming HIV infected within 2.56 and 1.7 years, respectively (Table 1).

Table 1. Risk profiles for baseline, 2x and 3x risk groups. In each case it is assumed that there are 0.2 sexual events per day.

This cursory analysis shows that, while a doubling or tripling of infection risk may not warrant concern given a single sex act, many such higher-risk sex acts do potentially increase overall infection risk significantly. The choice of parameters in this investigation suggest that subtle changes in risk at the per sex act level can make large differences to epidemic dynamics. This is because individuals in high risk groups tend to become infected faster, and in turn infect other people faster, thereby amplifying the growth of the epidemic. Those in low level risk groups may not infect other people during the course of their infection, while those in higher risk groups, are likely to infect one or more.

The above analysis of the problem indicates that health authorities should take small (2-4 fold) increases in HIV infection risk at the per-sex-act level seriously.

In contrast to classical analysis of individual-level data, which assesses associations between individual’s characteristics and an outcome of interest for that individual, network analysis offers a statistical framework to evaluate the importance of individual’s characteristics from other people to whom the individual is connected in the network, be it directly or indirectly through sexual relationships. Consequently, critical transmission pathways and key sub-populations that share network characteristics essential for propagation or disruption of the HIV transmission cycle may be identified. For example, De et al. discovered that individuals playing a central role in the spread of gonorrhoea do not necessarily have a large number of relationships, but have a high information centrality, i.e. a short average distance to all other people in the network (1).

As the full potential of network analysis for HIV epidemiology remains to be unlocked, so do the often powerful and refreshing implications of the results from HIV transmission and sexual network analyses: a more precise definition of key groups and transmission pathways could help to improve the effectiveness of HIV prevention interventions. Further, the efficiency with which information on HIV prevention and treatment diffuses in the population may be improved if the socially influential individuals and information flows are detected and prioritized. Both these considerations are increasingly relevant in light of the growing pressure on national and global funding for HIV prevention and treatment.

Finally, network analysis unifies individual-level and network-level information. Overall descriptions of the entire community or population can be generated from individual data; conversely, individual behaviour can be predicted or simulated from macroscopic attributes of the network.

A questionnaire survey exploring sexual histories is being conducted in the Cape Town area. Statistical analysis will extract the information of sexual networks and sexual behaviour from individual level data in three communities with a high burden of HIV. The statistical models will be used as input for SIMPACT, a software tool to simulate the spreading of HIV and impact of intervention under alternative sexual network structures and treatments strategies (2). Network analysis will play an important role in this study in two aspects. Firstly, it can help to understand and measure the patterns of sexual behaviour of South Africans. Furthermore, the influence of these patterns on the HIV epidemic and effectiveness of interventions can be simulated by SIMPACT. Secondly, network analysis and simulation can contribute in optimising impact of expansion of treatment by comparing potential outcomes of alternative expansion strategies.

We used an individual-based micro simulation model of serodiscordant heterosexual couples in South Africa to investigate these questions. The model included transmission of HIV-1, disease progression, use of ART, ageing, conception and pregnancies. The model was parameterised using data from the Partners in Prevention HSV/HIV Transmission study. Throughout the analysis two sets of behavioural assumptions were used: ‘Partners in Prevention couples’ (based on data from the Partners in Prevention study) and ‘More typical couples’ (assuming lower levels of condom use, more external partners and more unprotected sex with external partners). This was done in recognition of the fact that the couples observed in trials may or may not have similarities to the couples that would test together and be eligible for this kind of intervention in actual programmes.

First, we examined the impact of PrEP use by the uninfected partner before their infected partner started treatment. The highest impact, in terms of percentage of infections averted, is obtained when PrEP is used at all times by the uninfected partner after diagnosis of HIV in the infected partner, irrespective of ART use by the infected partner. Up to 59% of infections among couples could be averted with this strategy of PrEP use, assuming a PrEP effectiveness of 80% and the level of risk observed in Partners in Prevention couples. Strategies whereby PrEP is discontinued once the infected partner initiates ART result in a lower impact, with 49% of infections averted — but this was much more cost-effective. In fact, if PrEP is used in that way and the couples are at a high risk of becoming infected, using PrEP could reduce future costs as many individuals would then not become infected and require ART.

Next we compared using PrEP for the uninfected partner to earlier ART for the infected partner and attempted to identify the characteristics of PrEP that would be required for it to be at least as effective as providing earlier ART to the infected partner. For couples like those in the Partners in Prevention study, PrEP would be at least as cost-effective, at keeping couples ‘alive and HIV free at 50’ as earlier ART if its effectiveness was over 75% and the annual cost was less than 40% that of ART. In the trials, the upper estimates of effectiveness of oral PrEP did reach this level, and cost estimates do suggest PrEP would be substantially cheaper than ART, so this might be achievable, providing that couples using PrEP in real programs benefit from the same level of effectiveness as observed in the trials.

In summary, the evidence of efficacy of antiretroviral-based prevention might provide the best opportunity for reducing HIV infections. For sero-discordant couples, a key risk group for HIV transmission, the combination of both PrEP and ART may be a cost-effective way to reduce infections. However, there will certainly be many other considerations besides cost-effectiveness that inform decision-making for HIV treatment initiation and provision of PrEP in couples, including equitable access to medications and the preferences of the couples themselves.

SACEMA had the privilege of hosting an intense (Gates Foundation Funded) workshop of the international ‘HIV modelling consortium’, focusing on models of early treatment and their potential impact. Model scenarios were intriguing, debate was intense and constructive, and the prognosis was sobering. Treatment alone, at conceivable levels, given the current financial and infrastructural outlook, can curtail HIV incidence, but does not appear to open a road to HIV eradication. So we must accept that we must look to the ‘multiple prevention methods’ paradigm, new technologies, and cleverer healthcare systems if there is to be a future where HIV is a tolerable burden on those societies currently bearing the greatest impact.

An interesting theme worth watching, in the context of debate on interventions in the face of what will inevitably be a multi-generational epidemic, is the limits of orthodox discourse offered by various professional fraternities like ‘modellers’, ‘biostatisticians’, and ‘health economists’. The lengthy time scales, over which interventions and their consequences must play out, stretch all known, and perhaps all conceivable, formal methods to their limits, and demand that interpretation be a nuanced and careful technical execution. Should the model outputs be treated as crystal ball predictions, warnings, structural insights, or mere mental gymnastics. Perhaps it depends on the actual model in question, but formal predictions, in anything resembling the predictions which are the bread and butter of natural science, are hardly likely to be what epidemiological modelling is about.

On a narrower scope than steering us towards an HIV free generation, this edition of the SACEMA Quarterly epidemiological update offers, as usual, three perspectives on current challenges and recent progress. Carel Pretorius and Samuel Manda explore superficially very different, but, at a deeper level, analogous analytical problems relating to making sense of epidemiological ‘clusters’. It is tempting to attach interpretations to limited data which often does not support our pat conclusions, and we hope these pieces offer a useful view of on-going technical developments to support sound joining of the dots. A popular theme here at SACEMA, HIV incidence estimation, returns once again with a piece by Reshma Kassanjee, demonstrating how existing, previously neglected, specimens from the blood donation industry can be used to perform low cost, low risk, preliminary investigations into surveillance-application performance of new laboratory assays.

The first SACEMA Quarterly update of 2010 was devoted to the topic of HIV incidence (the rate of occurrence of new infections in a population). Incidence will always remain a crucial measure in epidemiology, providing a direct and current indication of the spread of disease. Prevalence (the fraction of the population with a condition at a point in time) is a metric more commonly measured, but is less informative as it emerges from historic incidence, demography and survival dynamics. South Africa has the largest HIV-positive population, exceeding 5 million individuals in 2011 (1). This year the government will launch its second 5-year National Strategic Plan (NSP) against HIV for 2012-2016. The headline goal of the first 5-year NSP (2007 – 2011) was to halve incidence, but it will be difficult to assess whether this goal has been achieved.

There are a number of approaches for estimating HIV incidence, with varying tractability, complexity and limitations. (i) In a prospective longitudinal study of a cohort of initially uninfected subjects, infection events are directly counted. However, such studies can be costly and prone to unrepresentative sampling. (ii) HIV prevalence, measured in sentinel or general populations, is often modelled to estimate incidence. For example, in age groups with low HIV mortality, differences in prevalence by age may be attributed to new infections. Alternatively, prevalence, measured at multiple time points in the past, could be used to estimate historic incidence. However, prevalence data has become increasingly complex to interpret as epidemics mature, and knowledge of post-infection mortality is often limited. (iii) Alternatively, back-calculation from AIDS cases involves estimating the historic HIV incidence that produced the observed AIDS incidence. However, this method provides little indication of recent incidence. An extension of this method utilises reported HIV diagnoses too. (iv) There are also a number of detailed ‘microscopic’ models, such as the UNAIDS Modes of Transmission model (see SACEMA Quarterly, March 2010 ), and dynamical models. These typically explicitly model the mechanisms of transmission of the virus through the population, requiring a number of quantitative assumptions.

Additionally, in recent years, there has been considerable interest in estimating HIV incidence from single cross-sectional surveys testing for ‘recent infection’ through laboratory-measured host or viral biomarkers (2). In a survey, the sizes of the HIV-negative, ‘recently infected’ and ‘non-recently infected’ populations can be measured, and incidence estimated using knowledge of the dynamics of the ‘recent infection’ biomarker (3,4,5).

Given the potential benefits arising from using single cross-sectional surveys for incidence estimation, this approach has been applied in numerous studies, and has caught the attention of prominent organisations worldwide. The World Health Organisation (WHO) Technical HIV Incidence Assay Working Group (HIVIWG) produced an extensive guide on the use of biomarkers for ‘recent infection’ for incidence estimation. The Centers for Disease Control and Prevention (CDC) continues to actively improve laboratory tests used to measure biomarkers that identify ‘recent infection’, and, earlier this year, supported the WHO’s efforts by hosting the latest HIVIWG meeting (August 2011). Notably, the Bill and Melinda Gates Foundation (BMGF) awarded the Health Protection Agency (HPA) a grant to assess, compare and optimise recent infection tests. The group working on this three-year project (2011-2013) is called CEPHIA, the Consortium for the Evaluation and Performance of HIV Incidence Assays, and comprises HPA, Blood Systems Research Institute (BSRI); University of California, San Francisco (UCSF); and SACEMA.

Two key obstacles to cross-sectional biomarker-based incidence surveillance remain: the (i) lack of standardisation of terminology and methodology, and (ii) poor characteristics, and characterisation, of currently available tests.

Characterising a Test for Recent Infection

Testing for recent infection for the purposes of incidence estimation differs, in key respects, from estimating how long each subject in a study has been infected. It is generally accepted that, in the context of incidence surveillance, there are two crucial characteristics of a test for recent infection (5,6).

Firstly, a mean duration of recent infection captures the average time spent ‘recently infected’. To produce reliable incidence estimates, the times that seroconverters spend ‘recently infected’ should be sufficiently large so that an adequate number of subjects are observed in this state in a survey of feasible sample size. For example, testing for the absence of HIV-antibodies amongst those with detectable HIV viral loads in principle identifies recent infection. However, the transiency of this state implies that very few ‘recently infected’ subjects would be found in a cross-sectional survey. Conversely, the ‘recent infection’ classification should not endure for extended periods of time post infection, as subjects who were infected long into the past will appear ‘recently infected’ in the survey, making incidence estimates less informative about current incidence.

Due to the substantial variability of the virus progression and immunological response, individual seroconverters can spend vastly different times classified as ‘recently infected’. To capture that some seroconverters are classified as ‘recently infected’ long after infection, a second characteristic, termed the false-recent rate, has been introduced. This measures the proportion of individuals that appear ‘recently infected’ a long time post infection. This proportion needs to be small to limit uncertainty in incidence estimates.

For a test for recent infection to be potentially useful for incidence surveillance, a mean duration of recent infection of 4 to 12 months and a small false-recent rate, less than 2%, are considered acceptable (6). Crucially, to produce robust incidence estimates, the characteristics of the recent infection test must be well-known (7).

Seroconverting Blood Donors as a Resource for Characterising

Recent Infection Tests In the past, estimation of recent infection test characteristics has relied on detailed longitudinal data. Specifically, data obtained from the regular follow-up of HIV-negative subjects, and then repeated testing for ‘recent infection’ amongst seroconverters, with small inter-test intervals, has been used. Such data is typically used to model the time each seroconverting subject spends ‘recently infected’, and this information is in turn used to estimate the average dynamic that is captured by the mean duration of recent infection. However, such data is expensive and logistically difficult to collect, and this has been an obstacle to the development of recent infection tests. Methods for using less well-characterised, but more easily captured, data could therefore greatly advance developments in this field. One such innovation is explored in the article ‘Seroconverting Blood Donors as a Resource for Characterising and Optimising Recent Infection Testing Algorithms’ (8), as briefly summarised below.

In the work, a readily-available source of specimens was identified, namely that of serocoverting blood donors. Utilising specimens from blood donors provides unique efficiencies as blood for transfusions is routinely collected and tested for HIV in most countries. In the study, repeat donors (in South Africa and the USA in the period 2001-2009) who were observed to seroconvert were tested for ‘recent infection’, using the specimens collected at the times of the first seropositive donations.

Data captured from such study designs has been overlooked in the past, because there is no follow-up of seroconverters and typically large intervals between HIV-tests (or donations). Such data provides little detail at an individual subject level. A method of estimation that draws meaningful information from this data, about average biomarker dynamics, without focusing on individuals, was employed. The approach is based on maximising the likelihood of the overall set of observed ‘recent’ and ‘non-recent’ classifications. The technical details of the method are provided in the abovementioned article (8). For example, if all inter-donation intervals were fixed at one year, then (for each donor equally likely to have been infected at any time in the year between his/her HIV-negative and HIV-positive donation) the overall collection of classifications provides information about the average dynamics of the biomarker for one year post infection. In the work, this principle is extended to allow for varying inter-donation intervals.

While the estimates of the test characteristics obtained using this method are likely not sufficiently robust for the purposes of incidence estimation, the work demonstrates an approach to perform preliminary characterisations of tests for recent infection, using a readily available source of specimens. More precise data could subsequently be used to better characterise only the most promising tests.

In conclusion, the cost and difficulty in collecting very detailed data describing the dynamics of a biomarker testing for ‘recent infection’ has been an obstacle to the characterisation of these tests, and hence incidence estimation. The innovative preliminary characterisation of recent infection tests, utilising more easily-sourced data, is therefore a promising development in the field. Using tests for recent infection to estimate times since infection for individual subjects is potentially of great interest in public health – this is a fundamentally different approach and would require an appropriately modified way to characterise tests for recent infection. Of interest in this discussion is the application of biomarkers testing for ‘recent infection’ for incidence surveillance. The ability to estimate incidence from a study performed at a single point in time offers great benefits, and therefore cross-sectional biomarker-based incidence estimation has drawn much interest in recent years. With the collective efforts of leading organisations and experts to address the remaining limitations in the field, there seems to lay a promise for many exciting developments in this area. Cross-sectional incidence estimation is certainly a topic that should be closely followed by epidemiologists, and could greatly advance population-level HIV incidence surveillance in years to come.  

The dataset consists of registered TB events, corresponding to the approximate time when a TB transmission event occurred (1). Clusters formed by TB strains isolates W451 and CC100 among HIV+ cases were particularly striking. These clusters may point to ongoing transmission, which could exasperate the already desperate situation in the township. A number of questions arise. The data show apparent clusters, but what are the properties of a typical cluster? Are the clusters we see the result of correlated infection events, e.g. an infection chain between HIV+ TB cases, or do they simply appear to be clustered when in reality there is no connection (correlation in statistical parlance) between them?

Mathematical theory of point processes

The best guide for this journey is the statistical theory of point processes. It helps us frame the questions we need to ask in order to interpret TB event data correctly. We developed a point process theory for TB events starting from the simple differential equation dynamical model we previously developed to understand the population level (macroscopic) aspects of TB in this community (2). To start building a point process theory for the TB model, we developed a dynamical description of all the random events that occur in the model, comprising birth, death, primary infection, re-infection and endogenous-activation events among susceptible, latently and actively infected sub-populations. To make this step analytically tractable we used only one HIV state in the model.

We then used van Kampen’s ‘population size’ expansion to derive a differential equation for the variances and co-variances of these random and fluctuating events (3,4). This is a so-called Fokker–Planck equation (FPE) and it describes the fluctuations as Gaussian noise around the equilibrium population level model. We solved the differential FPE with a standard ordinary differential equations (ODE) solver in Matlab, and checked the result against Gillespie’s stochastic simulation technique (5). Finally, we used the FPE to study the temporal aspects of TB clusters, and obtained an understanding of the timescale between active TB events.

Insights from point process theory

The first insight from the model applies to many deterministic models (see (2) for a summary)) used nowadays to model epidemics at the community level. We showed that fluctuations in the population variables (i.e. the variables that keep track of the number of susceptible, latently and actively infected individuals) become small relative to the size of the sub-populations when the population size is closer to 40,000. Given that many TB interventions and trials are often run and evaluated at a community level, we should expect a significant level of uncertainty in population-level estimates derived from macroscopic models. This uncertainty is seldom explicitly handled in the growing field of epidemic modeling, even for models applied to small communities.

The two-time correlation function

The next important insight from the model derives from the so-called two-time correlation function for events g2(t1,t2) (3, p. 41). It measures increased probability of observing an event of a particular type at time t2, given an event of a certain type at time t1. This is equivalent to measuring the degree to which the joint density of events at t1 and t2 is greater than the density at t1 and t2. This can be thought of as reflecting the causal influence of the first event on the second via both direct (e.g. infection, reactivation) and indirect routes (chains of such events): what is the increase in probability for an event to occur at t2 knowing that another event occurred at t1? Note that the events do not have to be of the same type.

Implications for TB control

Thus, even though chains of causation are not explicitly tracked in this type of framework, influence can be assessed at the statistical level by examining the correlations. This gives us a fairly detailed statistical description of clusters of events. For example, events that are separated by a timescale longer than the correlation timescale are unlikely to be part of the same transmission cluster. The method can therefore be used as the basis for understanding the temporal component of TB strain clusters, which are currently defined mostly in terms of DNA type, with geographical linkage, social interaction and other processes also playing a role.

Our analysis shows that endogenous activation events are correlated over long time scales, and are statistically likely to be part of short timescale clusters. This casts reasonable doubt around the presence of apparent clusters of isolates W451 and CC100 among HIV+ TB cases, and whether the data support the assumption of ongoing transmission of these strains. If the modeled intensities of active TB events are validated, then the model can be used to warn against spurious conclusions from measured clustered data. For example, a correlation effect may be incorrectly attributed to an infection trend, or even to a particular type of infection chain, while in reality it could be purely due to chance stemming from fluctuation. These observations have direct implication for TB control measures in the community: ongoing infection chains require more forceful intervention than reactivation events, which can be handled through standard TB control strategies.

Improvement and further work

Previous modeling work (6,7) has highlighted the potential for study time-windows and case-detection rates to bias interpretations of clustering statistics. Far less is known about how these may differ between data derived from HIV+ and HIV- TB cases. To do a complete analysis of a model with both HIV- and HIV+ TB cases is possible but cumbersome. We simply relabeled the HIV state in our model to HIV+, and changed all the dynamical parameters to ones corresponding to HIV+ individuals, who experience higher rates of primary infection, re-infection and endogenous reactivation. Higher rates of reactivation among HIV+ individuals may mean that their strains are a priori more likely to be drawn from the latent pool. Indeed, the analysis shows shorter correlation timescales among HIV+ individuals which suggests that a more stringent criterion of temporal linkage may be needed for cluster determination among HIV+ TB cases, compared with HIV- TB cases.

As more detailed data spanning a longer time interval become available it may be possible to evaluate if clustered active TB episodes are consistent with the dynamics of a closed community. If they are not, and therefore require exposure to external sources of infection to explain the observed clustering of TB events, it raises concerns for TB treatment programs. Treating TB cases only in a particular community will not reduce its TB burden: TB treatment programs would have to reach the wider community in order to be effective.

This analysis could find broad validation in epidemiological models where the transmission term is an assumed non-linear term, with few possibilities of validating the assumption against real data. The model is usually checked against aggregated population count data, which can be fit by many functional forms for the transmission term. Our approach of studying the underlying point process may shed light on whether a mass action model can produce the observed clustering of TB events. A model accounting for local contacts (household, schools) as well as global contacts (e.g. in the wider community) may be essential for modeling temporally clustered active TB events.

Note that the above is an abbreviated version of the following article: Pretorius C, Dodd P, Wood R. An investigation into the statistical properties of TB episodes in a South African community with high HIV prevalence. J Theor Biol. 2011;270(1):154-63. Link to article