Our results build upon the earlier finding in this cohort (9) which showed that uninfected infants of HIV-infected mothers with CD4 counts <350 were 2.9 times more likely to die by 4 months than infants of mothers with CD4 count >350. The current analysis in the same population found that after adjusting for socioeconomic status, children born to mothers with CD4 count <200 had 3.4 times the risk of death through 18 months as did infants born to mothers with high CD4 counts. A similar finding was observed in a cohort of uninfected children of HIV-infected mothers in Kenya, where children of mothers with a CD4 count <200 had 2.8 times the risk of death compared with all other CD4 counts.7
Our results differ from a pooled analysis which found a small, non-significant effect of a mother's CD4 count <200 compared with a CD4 count >500.23
They found a strong effect of maternal death on infant mortality and these variables likely both measure maternal disease stage; additionally, their results were adjusted for breastfeeding cessation which, according to our biologic model, would adjust away some of the effect of CD4 count.
We also found that breastfeeding cessation occurred earlier among immunocompromised mothers compared with healthier mothers near the time of the birth of their child. Our results agree with an earlier finding by Sedgh et al.24
in Tanzania where women with a CD4 count <200 breastfed on average 3.8 months less than women with a CD4 count ≥500 near the birth of her child. They are also consistent with data from Botswana which showed an increased risk of morbidity in children of HIV-infected mothers if the mother had stopped breastfeeding.5
Unlike previous studies, we assessed the mediating role of breastfeeding in the CD4 count–child mortality relationship. Our analysis supports the hypothesis that breastfeeding cessation accounts for some of the effect of maternal CD4 count on child death among HIV-exposed but uninfected infants. However, most of the CD4 count–child mortality association remained after adjusting for breastfeeding, suggesting other factors also explain the observed association. These may include increased exposure to infectious diseases, low birth weight and poor growth, or a general inability to care for the child.
The current study employed a marginal structural model to adjust for confounding by LBW as it is likely both a confounder and causal intermediate. When using a conventional Cox proportional hazards model, we found a substantially larger change in the estimate of effect of CD4 count on child mortality through 6 months when adjusting for LBW. This suggests that some of the effect of maternal CD4 count may be mediated through birth weight and that conventional approaches to adjusting for LBW will be biased.
While we identified no other study of breastfeeding cessation as a mediator of the CD4 count–infant mortality relationship, the reduction in effect of maternal CD4 count and child survival seen in one previous study after adjusting for breastfeeding cessation is consistent with breastfeeding cessation being a mediator.23
In their crude analyses, CD4 <200 was associated with 2.5-fold increased risk of child death compared with children of mothers with CD4 count >500. After adjusting for breastfeeding and other covariates their odds ratio was reduced to 1.7 (95% CI 0.9–3.2); this is in agreement with our finding that some of the effect is being mediated through breastfeeding cessation. However, as their analysis was not designed to assess the mediating role of breastfeeding it is not clear whether this change reflects adjustment for the mediator or adjustment for other confounders in their model.
Considerable debate still exists regarding the possibility and optimal approach for separating direct and indirect effects. The parsing of direct and indirect effects requires certain assumptions:19,25–28
there must be no residual confounding between the exposure and the outcome, or between the causal intermediate and the outcome, and there must be no interaction between the exposure and the intermediate. While our study was observational and therefore subject to confounding, we identified few confounders of the CD4 count–child mortality relationship and we adjusted for LBW, the only confounder identified for the breastfeeding cessation–child mortality relationship.
Some feel these assumptions are overly restrictive28,29
while others feel the conditions for deconstructing total effects rarely occur.27
argue that the no interaction requirement is unlikely to be satisfied. However, others25
note that, ‘in the presence of interaction, it is the potentially estimable parameter—the fraction of the exposure effect that could be eliminated by control of the intermediate—that will usually be the parameter of public health interest’. Thus, even if our assumption of no interaction between CD4 count and breastfeeding has been violated, our results still imply that a portion of the relative increase in child deaths through 6 months attributable to advanced maternal disease stage could be prevented by increasing breastfeeding duration.
This study should be considered in light of several limitations. First, there could have been residual confounding in our analysis. However, because maternal CD4 count was measured only during pregnancy, and most of our data was collected after the exposure, we had few candidate variables for adjustment. While it is possible that important residual confounding remained, as we could identify no confounders in this analysis other than low birth weight (either statistically or through substantive knowledge) we consider this scenario unlikely.
Second, time of breastfeeding cessation could have been misclassified. Because the misclassification is of the intermediate variable, if it is non-differential with respect to the exposure and the outcome, it would be expected to reduce the ability to adjust for the intermediate. This implies that any reductions seen in the CD4 child–mortality relationship after adjusting for breastfeeding cessation would not completely control for the intermediate, so that the indirect effect observed would be expected to be an underestimate.30
Thus our results likely represent an underestimation of breastfeeding as a causal intermediate.
Finally, we did not have data on cause of death for the current analysis, and therefore we cannot exclude the possibility that some of the child deaths were unrelated to breastfeeding (e.g. deaths due to injury). The impact of including these deaths is unclear, and future research on the topic should seek to refine the outcome.
In conclusion, we found that uninfected children born to HIV-infected mothers with more advanced immunosuppression were breastfed for shorter duration and had higher mortality though 18 months of life compared with children of mothers who were not immunocompromised. If confirmed, our results must be taken into account when deciding whether or not to recommend breastfeeding for infants of HIV-infected mothers. Our results suggest it is particularly important for mothers with low CD4 counts near delivery to persist in providing breast milk to their uninfected infants, though this risk must be balanced with the increased risk of HIV transmission with longer duration of breastfeeding. The more widespread availability of highly active antiretroviral therapy should also allow breastfeeding mothers the opportunity to continue to prolong breastfeeding without substantial increased risk of HIV transmission and therefore provide their uninfected infants the added protection and benefits of breast milk during the critical first months of life.31