All clinical trials with behavioral interventions are vulnerable to misleading results depending on the degree of adherence. Poor adherence may dilute the outcomes of a beneficial intervention leading to the erroneous conclusion of no benefit. However, if an intervention is harmful, then poor adherence may lead to the erroneous conclusion that an intervention is safe when it is not. The primary outcome of our trial was a combination of HIV infection or uninfected child death and early weaning affects each of these outcomes differently. The analysis we present here demonstrates that, after taking into account the dilution introduced by poor adherence, gains achieved in transmission reduction due to early weaning were offset by increases in mortality in the group overall. Hence the original conclusion of no benefit of early weaning is correct if no distinction is made among mothers at different stages of HIV progression.
Investigation into possible confounding led us to identify an interaction between severity of maternal HIV disease during pregnancy and effects of early weaning. This interaction has profound implications for infant feeding policies for HIV-infected women in low resource settings. If infants were born to mothers with advanced HIV disease, i.e. symptomatic and low CD4 counts, then early weaning resulted in fewer HIV infections or deaths. If infants were born to mothers with less advanced HIV disease, i.e. asymptomatic and high CD4 counts, then early weaning resulted in more adverse clinical outcomes. Since women with advanced disease should receive ART routinely as part of HIV treatment services 
, we anticipate that post-natal transmission should also be reduced. This has been shown in some demonstration projects 
. Therefore, if access to ART is in place, there is little justification for shortening the duration of breastfeeding among mothers with advanced disease. For women who do not yet need ART according to usual guidelines, our data indicate that risks associated with transmission are not large enough to outweigh the substantial increases in uninfected child mortality that result from prematurely truncating the usual duration of breastfeeding. For infants of mothers with less severe disease, weaning at 15 months or earlier is associated with a more than 3-fold increase in HIV infection or death. Even taking into account continuing risks of HIV transmission, HIV-free survival is better if breastfeeding continues for at least 16 months.
There are new data to support proof of concept that antiretroviral drugs given to the infant during breastfeeding can reduce postnatal transmission 
. Unfortunately, the studies investigating this have tested only short interventions 
premised on the assumption that infants of HIV-infected women could be safely weaned early. Studies are needed to evaluate interventions that extend over a full ordinary duration of breastfeeding.
A limitation of our study was that adherence with the intervention was incomplete. An advantage of intent-to-treat analysis of randomized controlled trials is that confounders associated with the practice of interest are less likely to explain associations. However, substantial lack of adherence with an intervention may bias associations towards the null, making intent-to-treat results difficult to interpret. We had not originally anticipated that early weaning would be so poorly accepted in our study population. The secondary analysis we present here is, like all analyses of epidemiological data, vulnerable to confounding and reverse causation. To address reverse causation, we reviewed the circumstances of all deaths, and weaning was only classified to have occurred if breastfeeding had ceased before the illness that preceded the child's death, and not as a result of that illness. We acknowledge that this cannot be foolproof but our approach avoids one of the methodological problems that has been previously identified in the field of breastfeeding research 
. We also examined a large number of confounders, including several known to be strong predictors of HIV transmission, specifically maternal CD4 count and viral load, and those known to be strong predictors of child mortality, namely birth weight and socioeconomic standing. None of these factors explained away the weaning associations. However, it is possible that other unmeasured confounders may account for our associations. The associations between weaning and adverse outcomes were consistent across the intervention and control groups. This makes it highly unlikely that social desirability tendencies or reporting biases (which would operate in different directions in the two groups) would explain the findings. Moreover, the benefits of continued breastfeeding in the women without advanced HIV disease are similar to those reported in non HIV infected populations 
Current WHO guidelines encourage weaning to occur only once affordable, feasible, acceptable, sustainable, and safe (AFASS) alternatives to breast milk are available 
. If guidelines are followed, HIV-infected women would only have initiated breastfeeding if AFASS criteria were not met during pregnancy. Since socioeconomic circumstances are unlikely to change between pregnancy and 6 months, it is unclear what this recommendation means for the appropriate age of weaning. Using HIV-free survival as the outcome, our data demonstrate that women from urban, resource-limited settings, who have high CD4 counts, place their infants at significantly increased risk if breastfeeding is stopped at any time before 16 months. There is a net benefit of continued breastfeeding for HIV-free survival of infants when maternal CD4 counts exceed about 300 cells/ul. For women with lower CD4 counts, initiating ART as expeditiously as possible during pregnancy should be an urgent priority, for their own health and to reduce transmission to the child.