Pregnancy during follow-up as exposed person-time. Incident pregnancies were identified from clinical 
records, as well as from records of antiretroviral drug regimens which list both pregnancy and “end of pregnancy” as reasons for regimen change.Figure 1. Cumulative risk of first incident pregnancy since HAART initiation, stratified by baseline age. doi:10.1371/journal.pone.0058117.gPregnancy and Clinical Response to HAARTTable 2. Estimated effect of pregnancy on time to death and alternate outcomes among 7,534 women initiating HAART in South Africa, 2004?011.Death Unadjusted Not 
pregnant Pregnant Weighted{ Not pregnant Pregnant Death or AIDS Stage 4 Unadjusted Not pregnant Pregnant Weighted{ Not pregnant Pregnant Death or AIDS Stage 3/4 Unadjusted Not pregnant Pregnant Weighted{ Not pregnant Pregnant Lost to follow-up Unadjusted Not pregnant Pregnant Weighted{ Not pregnant PregnantNo. of event{order AKT inhibitor 2 Person-months of follow-upHR95 CL614220,093 29,1. 0.67 0.43, 1.456181,558 24,1. 0.84 0.44, 1.868209,694 27,1. 0.80 0.56, 1.677172,832 22,1. 0.87 0.51, 1.1,105199,123 25,1. 0.92 0.67, 1.876163,912 21,1. 1.13 0.72, 1.1,796220,093 29,1. 0.70 0.59, 0.1,349181,558 24,1. 0.62 0.51, 0.HR, hazard ratio; CL, confidence limit. Weighted models SPDB supplier accounted for age, employment status, active tuberculosis at study entry, calendar date at entry, WHO stage, and baseline and time-updated measures of weight, body mass index, hemoglobin, CD4 count and percent, adherence, and current drug regimen. { Difference from unadjusted model due to missing data in any variable; only complete observations get weights. doi:10.1371/journal.pone.0058117.t{The main outcome in this study was death; we considered a secondary outcome of death or a new AIDS defining event [33]. Deaths were 1531364 obtained from the clinic database, and from the national death registry [34]. AIDS-defining conditions were obtained from clinical records; please see File S1 for details of this definition. Lost to follow-up was defined as not being seen in clinic for six months, no evidence of clinic transfer, and no record of death in the national death registry.Statistical AnalysisBaseline characteristics of women were described using simple statistics, including chi-square tests for categorical variables and Wilcoxon rank sum tests for continuous variables. As noted above the 10457188 main analysis focused on the effect of incident pregnancy on time to death. A key concern of this analysis was the possibility of time-varying confounding affected by prior exposure. [35,36] This situation might manifest as follows: the effect of pregnancy on time to death is likely confounded by CD4 count (as a proxy for underlying immune status); but CD4 count might itself be affected by pregnancy. Situations in which time-varying confounding affected by prior exposure is likely are best analyzed using inverse probability weights to estimate marginal structural Cox proportional hazards models [35,36] and confounding-adjusted extended Kaplan-Meier curves [37]. Inverse probability weights account for bias due to both confounding [38] and drop-out. [35] In the main analysis, we censored inverse probability weights at the 0.1st and 99.9th percentiles to reduce the overall variance of our estimates and to prevent single individuals who are exposed against expectations (e.g., a very immunosuppressed, older woman who becomes pregnant) to exert undue influence on results of analysis. [39]. In all multivariable analyses, we considered the follo.Pregnancy during follow-up as exposed person-time. Incident pregnancies were identified from clinical records, as well as from records of antiretroviral drug regimens which list both pregnancy and “end of pregnancy” as reasons for regimen change.Figure 1. Cumulative risk of first incident pregnancy since HAART initiation, stratified by baseline age. doi:10.1371/journal.pone.0058117.gPregnancy and Clinical Response to HAARTTable 2. Estimated effect of pregnancy on time to death and alternate outcomes among 7,534 women initiating HAART in South Africa, 2004?011.Death Unadjusted Not pregnant Pregnant Weighted{ Not pregnant Pregnant Death or AIDS Stage 4 Unadjusted Not pregnant Pregnant Weighted{ Not pregnant Pregnant Death or AIDS Stage 3/4 Unadjusted Not pregnant Pregnant Weighted{ Not pregnant Pregnant Lost to follow-up Unadjusted Not pregnant Pregnant Weighted{ Not pregnant PregnantNo. of event{Person-months of follow-upHR95 CL614220,093 29,1. 0.67 0.43, 1.456181,558 24,1. 0.84 0.44, 1.868209,694 27,1. 0.80 0.56, 1.677172,832 22,1. 0.87 0.51, 1.1,105199,123 25,1. 0.92 0.67, 1.876163,912 21,1. 1.13 0.72, 1.1,796220,093 29,1. 0.70 0.59, 0.1,349181,558 24,1. 0.62 0.51, 0.HR, hazard ratio; CL, confidence limit. Weighted models accounted for age, employment status, active tuberculosis at study entry, calendar date at entry, WHO stage, and baseline and time-updated measures of weight, body mass index, hemoglobin, CD4 count and percent, adherence, and current drug regimen. { Difference from unadjusted model due to missing data in any variable; only complete observations get weights. doi:10.1371/journal.pone.0058117.t{The main outcome in this study was death; we considered a secondary outcome of death or a new AIDS defining event [33]. Deaths were 1531364 obtained from the clinic database, and from the national death registry [34]. AIDS-defining conditions were obtained from clinical records; please see File S1 for details of this definition. Lost to follow-up was defined as not being seen in clinic for six months, no evidence of clinic transfer, and no record of death in the national death registry.Statistical AnalysisBaseline characteristics of women were described using simple statistics, including chi-square tests for categorical variables and Wilcoxon rank sum tests for continuous variables. As noted above the 10457188 main analysis focused on the effect of incident pregnancy on time to death. A key concern of this analysis was the possibility of time-varying confounding affected by prior exposure. [35,36] This situation might manifest as follows: the effect of pregnancy on time to death is likely confounded by CD4 count (as a proxy for underlying immune status); but CD4 count might itself be affected by pregnancy. Situations in which time-varying confounding affected by prior exposure is likely are best analyzed using inverse probability weights to estimate marginal structural Cox proportional hazards models [35,36] and confounding-adjusted extended Kaplan-Meier curves [37]. Inverse probability weights account for bias due to both confounding [38] and drop-out. [35] In the main analysis, we censored inverse probability weights at the 0.1st and 99.9th percentiles to reduce the overall variance of our estimates and to prevent single individuals who are exposed against expectations (e.g., a very immunosuppressed, older woman who becomes pregnant) to exert undue influence on results of analysis. [39]. In all multivariable analyses, we considered the follo.
