Our primary finding is that circumcision alone will not be the ‘silver bullet’ that halts the HIV epidemic. A more likely scenario is for incidence to be eventually reduced by 25–35% if high coverage levels are achieved (). It has previously been suggested that the reduction in rate of HIV transmission from female-to-circumcised-male of 60% is comparable to a vaccine delivered to both men and women of about 37% efficacy, if all men are circumcised 
. Our model is good agreement with that finding, but although acceptability of male circumcision has been reported at promisingly high levels (circa
50% in many settings 
), complete coverage seems implausible. With lower coverage, a weaker net effect is projected ().
The indirect benefit of circumcision interventions to women (and uncircumcised men) is mediated by reduction in HIV prevalence among their circumcised male sexual partners (and partners' partners). It is slower to emerge because the long survival time with HIV means that prevalence declines gradually. Furthermore, since it comes via the sex partner network, its extent is extremely difficult to predict because sexual behaviour is multi-faceted 
, sometimes incompletely reported 
and because the pattern of transmission is also strongly linked to higher-order sexual network properties 
. Some simple models may fail to capture the full extent of this indirect effect and under-estimate the total impact of the intervention 
. Similarly, analyses that measure the impact of circumcision over only short periods (<10 years) will not quantify the full benefit of the intervention 
. In addition, because of the long term consequences of the intervention it is very important that the value of future benefits is appropriately discounted in economic analyses. For these reasons, other studies have shown that when a short-term time-frame is considered, quicker scale-up of services can substantially increase the overall impact 
The impact of any intervention depends on the existing patterns of risk and transmission in the population: the epidemiological context 
. Epidemics are sustained if the chain of transmission (one individual infecting another) is maintained. Generally, in communities with low risk, that chain is fragile and may be broken by small biological or behavioural changes. If there is more risk, the same changes have less impact because the chain is still maintained. However, the smaller absolute number of infections means that, in such settings, more operations may be required to achieve the same number of infections averted. Thus, in general, the potential proportional impact of interventions on the epidemic is greater in low-prevalence, low-circumcision groups. The lowest cost per infection averted, in contrast, will be achieved in higher prevalence communities. However, since many alternative patterns of risk can lead to the same endemic prevalence level, it not possible to accurately judge the impact of the intervention using only that information. For instance, “low” prevalence in a country can signify either a core of high-risk behaviour with the rest at no risk of infection (where the intervention could have little impact), or a more even distribution of moderate risk throughout the population (where the intervention could substantially contribute to arresting transmission).
In combination with other behavioural changes, the impact of circumcision interventions could be much greater. At the individual-level, men that protect themselves with condoms will get a disproportionately greater protective benefit from circumcision, and at the population-level, synergies between interventions will amplify the reduction in incidence. To avoid wasting resources and a unique opportunity, circumcision programmes must be accompanied by a renewed and vigorous focus on behaviour change 
. Circumcision programmes will also operate well alongside ART programmes. The modest reduction in new infections due to ART will be supplemented by reductions due to circumcision. That will lead to a reduced demand for ART in the future and, in the meantime, deaths due to AIDS will fall substantially.
Risk compensation could dent the impact of the intervention, so it will be especially important for safe-sex messages to be reinforced for men being circumcised. Increased risk behaviour could undermine derived benefits for women especially, but net increases in incidence (among women or the population overall) are only associated with very great increases in risk. Data from the three randomised trials 
and another cohort study 
did not find evidence for such large changes in risk following circumcision. Being able to avoid using condoms or having more sexual partners are not among the reported reasons for getting circumcised 
The conflicting evidence on the benefits of circumcising infected men 
, including the chance that transmission is increased if men resume sex before the wound has healed 
, and our modelling results leads to interesting ethical dilemma. There is clear advantage in circumcising infected men if the operation does reduce the chance of male-to-female transmission, but even if it does not and transmission is greatly increased during the healing period, the impact of the intervention for the population is not considerably reduced. However, individual specific women may be placed at greater risk. This has to be considered against the potential reduction in uptake if HIV-testing is a pre-requisite for being circumcised (necessary to avoid circumcising any infected men). On balance, in the interests of doing no harm, it is likely that the protection of the individual will outweigh the protection for the population. However, our modelling shows that this is borne out of a concern for the individual not the population.
We have explored the sensitive of our findings to the parameters specifying the pattern of heterosexual HIV transmission and the biological effect of circumcision and we expect that our conclusions will be applicable generally to the mature generalised epidemics of southern Africa. The precise impact of interventions will be determined by many local factors, including the epidemiological context, the level and pattern of uptake of the intervention, the biological effect of circumcision and the degree of risk compensation 
. It will be important for quantitative projections to be tailored to the local situation and to incorporate as much data as possible on the historic epidemic trends and sexual behaviour. We also recognise that the impact of the interventions will be lower if the biological effect of circumcision is less than was found in the carefully conducted and well managed trials 
It is important in modelling work to establish the influence of model structure on the results – that is, whether the conclusions drawn are linked to the formulation of a particular model. Our findings are in close quantitative agreement with different types of model that have focussed on other settings and employed alternative analytic techniques 
. Future work will aim to identify ways in which the impact of the intervention could be maximised, quantify uncertainty in projections and explore different techniques for predicting the impact of interventions, from micro-simulation to tractable analysis 
. Mathematical modelling must build upon the gold-standard evidence from the randomised controlled trials to provide both qualitative understanding and detailed quantitative predictions to support the decision-making processes that are now underway.