In this observational study of HIV-positive women initiating HAART in South Africa, we found that pregnancy was not associated with an increased hazard and risk of death nor with increased hazard of combined outcome of AIDS or death over a substantial period of follow-up (). These results were generally supported by sensitivity analysis. While we estimated hazard ratios below or approximately at the null for the effect of pregnancy on death or AIDS, it seems unlikely that pregnancy is truly protective against death; rather, keeping in mind substantive considerations, we suspect that these results reflect a null effect of pregnancy on risk of clinical response to HAART, rather than a protective effect. Put another way, we believe that these results argue simply that pregnancy does not increase overall risk of death in this setting.
The clinical results largely cohere with reports from the United States 
and South Africa 
, although both of those reports have some limitations. The US results are difficult to interpret due to methodological concerns 
, while the South African study concentrated on prevalent pregnancy rather than incident pregnancy, an approach which may have significant limitations 
. These results point in the opposite direction of previous results from this database showing that incident pregnancy may mildly accelerate rates of virologic failure 
, but are in line with another study of pregnancy and virologic failure 
. As this suggests, much remains unknown about the impact of incident pregnancy on response to HAART (if any): more work is necessary to understand this relationship.
In contrast, we found a robust association of pregnancy and reduced hazard of drop-out from the cohort (HR
0.62, 95% CL 0.51, 0.75). Unlike the clinical responses, the observed effect of pregnancy on drop-out may plausibly represent a protective effect of pregnancy. Such a protective effect might be observed if, for example, clinical providers were emphasizing the need to stay in care to new or expectant mothers, to protect the health of a newborn both directly (e.g., by preventing transmission) and indirectly (by preserving maternal health and thus enabling better care for the child). Of note, this finding stands in stark contrast to the association of prevalent pregnancy (at HAART initiation) with increased rates of lost-to-follow-up 
; however, some of that increased rate may be due to missed transfer rather than drop-out.
We saw little difference in crude drug adherence between non-pregnant and pregnant person-time, where we might expect to see higher adherence among pregnant women 
(although not necessarily 
). Some of this difference may be due to the fact that prior studies generally separated the pregnant and postpartum periods (and saw differences between them) 
, whereas we did not separate these two periods (see Methods
). As well, many existing studies have focused exclusively on prevalent, rather than incident, pregnancy: as prevalent and incident pregnancy appear dissimilar in their association with virologic failure rates 
, we might likewise expect them to be dissimilar in their association with adherence. We note that a recent large meta-analysis found that only 74% of pregnant women (and still fewer postpartum women) had optimal adherence to HAART 
, underlining the urgent need to identify HIV-positive pregnant women at risk of low- or non-adherence 
Beyond large sample size and a long follow-up period which included over 20,000 person-years of follow-up, there were several key strengths of this study. Data were collected prospectively in a previously validated clinic database. 
Issues of time-varying confounding affected by prior treatment were dealt with appropriately 
, using marginal structural Cox proportional hazards models 
and weighted Kaplan-Meier curves. 
Outcome collection was aided by the availability of the national death registry 
, and results were similar when restricting analysis to those patients who gave us a valid medical identification number (and thus whose vital status should have been captured in the registry). Finally, previous reports from this cohort 
suggest that results from the Themba Lethu Clinic are generally comparable to other large cohorts in sub-Saharan Africa; thus, we believe that the present results will have good generalizability to other urban HAART cohorts in sub-Saharan Africa.
The main limitation of this study was that, despite the large numbers of participants (7,534) and person-time (20,813 person-years), there were only very few exposed person-events (21 deaths in women who had experienced pregnancy), which limited power. However, one can also view this small number of exposed person-events as a result in itself: one which suggests that while it is not impossible that pregnancy during HAART is of clinical concern, it is unlikely to constitute a major overlooked public health issue in this context. One reason for these low numbers was that some number of pregnancies may have gone undetected due to miscarriage or early pregnancy loss. If such pregnancies were numerous, and strongly associated with both death, these results could be misleading: a more cautious interpretation of these findings, that recognized incident pregnancy was not associated with increased hazard of death or AIDS, may therefore be advised. That said, sensitivity analyses (not shown) suggest that unrecognized pregnancies would have to be very strongly associated with death (a much stronger association than was observed among recognized pregnancies) in order to substantively change our overall conclusions.
Other limitations of this study should be noted. We analyzed observational data from a clinical database, and thus uncontrolled confounding (in particular, by viral load, which was often missing and so was not controlled in main analysis) remains a possible threat to the validity of this study. Likewise, misclassification of exposure, outcome, and other factors cannot be ruled out. However, our sensitivity analyses support the results of our main analysis, and as noted above the TLC database has been previously validated for accuracy. 
At the same time, we note that observational studies are often more generalizable than randomized controlled trials, which typically have stringent inclusion and exclusion criteria. Moreover, as an exposure pregnancy cannot be randomized either ethically or practically; thus, robustly analyzed longitudinal observational data is by necessity the gold standard for research that considers pregnancy as an exposure.
Finally, there were substantial missing data in the main analyses. However, sensitivity analyses, including multiple imputation of missing values, suggests strongly that our complete case analysis is not biased in such a way as to alter the overall qualitative conclusions of this work. That said, we are missing information on precise causes of death, as well as on birth complications: efforts are ongoing to obtain the latter data.
In this large study of clinical cohort data, we found that recognized pregnancy after HAART initiation was not associated with increased relative hazard or absolute risks of death or AIDS, and may have decreased risk of drop-out. Nonetheless, this study reconfirms that incident pregnancy is common after initiation of HAART among younger women 
, and emphasizes the need for better integration of reproductive healthcare services and standard clinical care for HIV-positive women in sub-Saharan Africa.