The IeDEA-SA collaboration study was approved by the University of Cape Town and University of Bern human research ethics committees. The requirement for informed consent was waived, as only anonymized data that were already collected as part of routine monitoring contributed to the collaborative dataset.

This study is based on HIV treatment cohorts in southern Africa that took part in the IeDEA-SA collaboration (http://www.iedea-sa.org/). The collaboration has been described in detail elsewhere [17],[18]. In brief, data were collected at each site as part of routine monitoring and were transferred to data centres at the University of Cape Town, South Africa, or the University of Bern, Switzerland, at regular intervals. All programmes obtained ethical approval from the relevant local institutions before contributing anonymized patient data to IeDEA-SA. The present analysis is restricted to eight cohorts that routinely captured both pre-ART data and post-ART data of HIV-infected children: Lighthouse Trust Clinic in Malawi; Newlands Clinic in Zimbabwe; and McCord Hospital, Harriet Shezi Children's Clinic, Gugulethu Community Health Centre, Khayelitsha ART Programme, Hlabisa HIV Treatment and Care Programme, and Rahima Moosa Mother and Child Hospital in South Africa. All clinics except McCord Hospital were public health sector programmes with research support providing primary care to urban populations, but rural populations were included in the Hlabisa HIV Treatment and Care Programme [18]. At Harriet Shezi Children's Clinic and Rahima Moosa Mother and Child Hospital, secondary and tertiary care are available in addition to primary care.

In South Africa, the recommended first-line regimen was two nucleoside reverse transcriptase inhibitors plus either lopinavir/ritonavir (children <3 y of age or <10 kg) or efavirenz (children >3 y of age and >10 kg). In Malawi and Zimbabwe the recommended first-line regimen for all children was two nucleoside reverse transcriptase inhibitors plus nevirapine. The nucleoside reverse transcriptase inhibitors used were mainly stavudine or zidovudine plus lamivudine.

All children who were at least 24 mo of age at enrolment into HIV care, enrolled between 1 January 2000 and 30 June 2012, and were not older than 60 mo were included. Children who were not ART-naïve at their time of entry into the HIV care programmes or who had no follow-up data were excluded from the analysis. The detailed process of selection of children eligible for this analysis is provided in Figure 1.

Baseline characteristics included CD4 count, CD4%, weight-for-age z-score (WAZ), and height-for-age z-score (HAZ), as well as the child's sex and age at first visit. All z-scores were calculated using WHO standards [19]. Follow-up data consisted of CD4 count, CD4%, and WAZ, evaluated 1 mo after the first clinic visit and in 3-mo intervals starting 3 mo after the first visit, for up to 36 mo in care. If two or more observations per interval were available, the one closest to the middle of the interval was used. The date of ART initiation (if applicable) was captured as well. If there were no data for a child in a particular time interval, data were defined to be missing. The outcome was mortality and was ascertained by the individual cohorts. Children were defined as being lost to follow-up (LTFU) and were censored if they had no contact with their health care facility for 9 mo before database closure (visits are scheduled up to 6 mo apart and clinic visits being delayed for up to 3 mo are possible).

To account for missing follow-up data for CD4 count, CD4%, and WAZ, a child's last observation was carried forward for a maximum of 9 mo. For the remaining missing follow-up data, as well as missing baseline data for CD4 count, CD4%, WAZ, and HAZ, longitudinal multiple imputation by means of the expectation maximisation bootstrap algorithm [20] was used to create ten imputed datasets. Imputations were utilised only after 9 mo without any visit data, as from there on it is plausible that CD4/WAZ measurements that determine treatment assignment were taken (and are thus needed to adjust for time-dependent confounding) but not electronically recorded, probably because of administrative or clerical errors. The missing at random assumption is therefore likely to hold. The imputation model included all measured baseline and follow-up variables, mortality, follow-up time, a variable indicating which observations were carried forward, and the region (Zimbabwe, Malawi, South Africa urban, South Africa rural). The longitudinal structure of the data was explicitly considered in the expectation maximisation bootstrap algorithm, nonlinear time trends were allowed, and lag and lead variables of CD4 count, CD4%, and HAZ were added to the imputation model. All results of the multivariate analysis are based on the imputed datasets and were combined using Rubin's rules [21].

Baseline characteristics were described using proportions or medians (reported with first; third quartile) overall and for different ART initiation thresholds. Follow-up data were described at 12 mo and 24 mo after the first clinic visit.

We then estimated mortality over 3 y of follow-up for ART initiation at different CD4 thresholds using causal modelling (g-computation) [13], adjusting for time-dependent confounding of CD4%, CD4 count, and WAZ, with 95% bootstrap confidence intervals based on 200 bootstrap samples. WAZ served as a proxy variable for WHO stage, which was not routinely available in our data: many clinical events that define advanced clinical WHO stage, and thus disease progression and severity, relate to a child's WAZ—for example, malnutrition, tuberculosis, persistent diarrhoea, severe wasting, recurrent bacterial pneumonia, recurrent bacterial infection, candidiasis, and chronic lung disease [22]–[24]. WAZ was therefore expected to approximately adjust for the time-dependent confounding induced by stage-defining severe clinical events. The analysis was also adjusted with respect to baseline CD4 count, CD4%, WAZ, HAZ, sex, age, and region.

The primary comparison we considered was (i) giving ART immediately irrespective of CD4 criteria or (ii) deferring ART until a child's CD4 cell count falls below 750 cells/mm³ or CD4% falls below 25%; further strategies we considered were (iii) deferring ART until the threshold of CD4 count <500 cells/mm³ or CD4% <20% is reached, (iv) deferring ART until the threshold of CD4 count <250 cells/mm³ or CD4% <15% is reached, and (v) no ART is given.

The details of our g-computation algorithm are contained in Box 1. Briefly, the observed associations in the data were used to simulate and evaluate counterfactual datasets that would have been observed under each of the above initiation rules. This mimics an RCT that compares the effect of the different initiation rules. These datasets were standardised with respect to the time-dependent confounders (CD4 count, CD4%, WAZ) that influence both mortality and ART initiation in children [25]. To implement this, we first modelled the association of these time-dependent confounders at all follow-up visits (at time t = 1, 3, 6,… mo) with disease progression history (follow-up data at time t−1), baseline characteristics (CD4 count, CD4%, WAZ), demographics (age, sex, region), and ART (at time t−1, included under an intention to treat assumption) using additive linear models. We used additive logistic regression to model the association of death with the same variables at each follow-up interval. For all these models nonlinear relationships and interactions with baseline characteristics were explored and included if they improved the generalised cross-validation score. In the second step we used the observed baseline data and predictions of the fitted models under a specific treatment initiation rule to simulate the time-dependent confounders and the outcome “forward in time”. These two steps were repeated for all the above specified ART initiation rules, and mortality in the simulated datasets was evaluated thereafter. All these steps were repeated for all bootstrap samples in each imputed dataset and combined appropriately.

We further conducted a sensitivity analysis, linking data from three of the South African cohorts (Hlabisa HIV Treatment and Care Programme, Khayelitsha ART Programme, McCord Hospital) with the national vital registry to obtain the vital status of children LTFU. We then imputed remaining events and survival times for children LTFU and unascertained [26] and estimated mortality using the g-computation algorithm described above.

We also calculated the percentage of children who presented with CD4 count above the threshold of 750 cells/mm³ or CD4% above 25% and estimated the probability of falling below one of these thresholds during the 3 y of follow-up using the Kaplan-Meier estimator. This analysis was restricted to pre-ART data, and we censored both children who died and children who were LTFU. The same probability was also estimated by means of cumulative incidence functions treating ART initiation, death, and LTFU as competing events. All analyses were conducted using R 2.15 [27].

Our results indicate no mortality difference for up to 3 y between immediate ART initiation irrespective of CD4 values and deferred ART initiation using a CD4 threshold of CD4 count <750 cells/mm³ or CD4% <25%, though point estimates for mortality were higher when ART was initiated at lower CD4 values (or not at all).

To our knowledge, this is the first attempt to estimate the effect on mortality of starting ART at different CD4 values in children aged 2–5 y. Both RCTs and previous causal modelling studies of observational data included a wider age range of children, with relatively small numbers of children in the 2–5-y age group [6],[9],[28],[29]. Although post hoc subgroup analyses of the RCT data restricted to this age range have been conducted, results were inconclusive [30]. Moreover, the observational studies had few children without either clinically or immunologically severe disease and did not compare different CD4 thresholds but rather concentrated on the effect of ART itself [28],[29].

Our analysis included a large dataset of nearly 3,000 children with relatively long follow-up from three resource-limited countries with high HIV burden. Because of the good availability of CD4 and WAZ data, which is unusual in this setting, as well as the use of g-computation, we were able to adjust for measured time-dependent confounding of CD4 count, CD4%, and WAZ. In addition, we accounted for missing data using multiple imputation.

The use of a causal modelling technique was essential to accurately estimate mortality associated with starting ART at different CD4 values in the presence of time-dependent confounding by indication, whereby the illest children have the greatest risk of mortality and are also most likely to be started on ART. There was clear evidence for time-dependent confounding: a naïve analysis of the data simply adjusting for demographic and disease severity characteristics at enrolment showed that ART initiation was associated with higher mortality than no ART initiation (hazard ratio = 1.80; 95% CI: 1.03–3.15). Other causal modelling techniques such as inverse probability of treatment weighting, g-estimation, or targeted maximum likelihood estimation could also have been used to derive results that adjust for time-dependent confounding, though g-computation is probably the most intuitive when dealing with dynamic treatment comparisons [13].

Although we used causal modelling, this analysis is based on observational data, and the results may be affected by residual confounding. For example, WHO clinical stage data were not routinely available throughout follow-up, and so we may not have adequately adjusted for the effect of clinical disease severity. We addressed this possibility by adjusting for WAZ, which is a surrogate of disease severity and should at least partly capture staging information [23]. Some stage-defining events such as HIV encephalopathy or HIV-associated nephropathy may, however, not be reflected by WAZ.

When interpreting estimates of mortality associated with starting ART at high CD4 values, it should be noted that relatively few children initiated ART before their CD4 count or CD4% declined. Since country and WHO guidelines at the time did not recommend ART initiation at high CD4 values unless children had WHO clinical stage 3 or 4 disease, it is plausible that these children may have initiated therapy on the basis of advanced clinical disease. The median WAZ in children starting ART with CD4 >750 cells/mm³ or CD4% >25% was −1.42 (Table 1), indicating that a proportion of these children had clinically severe disease. Our analysis may therefore underestimate mortality reductions associated with starting ART at high CD4 values in a child without symptomatic disease. Nevertheless, our mortality estimates associated with immediate ART initiation or starting ART at a high CD4 value are consistent with the probability of experiencing AIDS or death found in both the immediate and deferred ART arms of the PREDICT trial [9]. It would be useful to repeat this analysis using data from other settings, for example, in high-income settings where more children start ART earlier in the course of their disease.

Further limitations of our study include not examining outcomes beyond 3 y of follow-up. This limitation is especially relevant with respect to evaluating toxicity risk and the effect of development of resistance in a context of limited second-line options. In addition, children included in our study were mostly from urban settings with facilities for electronic data collection in southern Africa, especially from South Africa. Generalisability to rural areas and sub-Saharan African countries with less well-resourced health care systems may be limited. In addition, country-specific differences regarding drug regimens, public health programmes, and breastfeeding [31] and its implications for the timing of HIV infection may contribute to both mortality figures and the success of different treatment strategies.

Our analysis was specifically targeted at examining the WHO 2010 guidelines for ART initiation for children aged 2–5 y and how mortality would be affected by the WHO 2013 guideline recommendations. There are, however, other interesting comparisons that remain unexplored: for example, we did not include children between 1 and 2 y of age. There are currently no data that show mortality benefits of early ART in this age group [3]; a comparison with the age groups in this analysis would therefore be useful.

The absolute mortality estimates for all ART initiation strategies in our study are relatively low [10],[28]. This finding may partly be explained by the analysis being restricted to children aged at least 2 y at their first visit to an HIV care and treatment facility. Our data therefore comprise a cohort of survivors who have not died and have not been referred for HIV care or ART prior to reaching 2 y of age, and so may have less rapidly progressing disease [10],[11],[32]. However, it is precisely in this group of children that the optimal timing of ART initiation needs to be determined, as WHO 2010 guidelines already recommend immediate ART in children presenting at less than 2 y of age [16]. Moreover, our mortality estimates are comparable to the results of the PREDICT trial, which reported a 2.1% and 1.3% risk of death/AIDS events after 144 wk in the early and deferred ART arms, respectively [9].

We may still have underestimated mortality. Children with no visits after enrolment were excluded from the analysis, as at least one follow-up visit is needed in order for children to have the possibility to initiate ART and thus allow the implementation of our methodology. The excluded children had higher CD4 counts (median [first; third quartile]: 694 cells/mm³ [382; 1,073]) and CD4% values (median [first; third quartile]: 21% [13%; 28%]) but lower WAZ values (median [first; third quartile]: −1.9 [−0.7; −3.3]), and their mortality may be different from those included in the analysis.

In addition, among the 2,934 children included in the analysis, those LTFU were censored at their last visit, and mortality may also be higher among those LTFU compared to those remaining in care [18],[33],[34]. To address this potential problem we conducted a sensitivity analysis, linking data from three of the South African cohorts with the national vital registry and imputing survival times and events of children LTFU and unascertained with multiple imputation as described previously [26]. This analysis confirmed that mortality might be higher than reported, though the relative differences between the strategies in the sensitivity analysis after 3 y of follow-up remained similar to those in our initial results ignoring LTFU that yielded much lower mortality estimates. It is worth mentioning that our estimate of mortality for children receiving no ART (7.8% [95% CI: 5.1%–10.4%]) is consistent with estimates of mortality from previous studies among children in this age group in the absence of ART [10]. However, the linkage information we had was sparse, the sample size was small, and data were available only from South Africa, so mortality estimates derived after linking data with the death registry should be interpreted with caution.

The WHO guidelines changed in 2013 and now recommend ART initiation in all children less than 5 y of age irrespective of clinical or immunological disease severity. Our results suggest that this change in ART initiation criteria would neither reduce nor increase mortality. There may, however, be other considerations regarding the optimal timing of ART. For example, while the PREDICT trial showed that the immune response to ART was equivalent in both groups, children in the deferred group spent more time at a lower CD4 count, which may result in more chronic inflammation and a higher risk of clinical events. This consideration is especially important for the older children in our age group, as their thymic responses are not as good as in children under 2 y [35]. However, there was no difference in the number of CDC category B or C events between the two arms in the PREDICT trial [8],[9]. The long-term effects of earlier ART initiation on immune function, growth, neurological development, toxicity, and development of resistance remain to be explored. In addition, maintaining long-term adherence can be particularly challenging in young children because of the need for a caregiver to administer medication and possible caregiver changes, lack of child-friendly formulations, and the need for frequent dose changes in growing children.

Programmatically it may be simpler to have a single treatment initiation approach for all children <5 y of age that does not require regular pre-ART follow-up and CD4 measurement, especially in settings where access to CD4 measurement is limited [36]. Coordination of recommended regimen choices with age thresholds for starting ART would also ease implementation. In our study only about 11% of children aged 2–5 y had CD4 count >750 cells/mm³ and CD4% >25% at presentation, and we estimated 72.2% (see Figure 3A) to have progressed below these values within 3 y, though the proportion might be even larger during early follow-up, given that children who were censored because they were LTFU or started ART may have had faster disease progression than others. Deferring treatment until CD4 count drops below 750 cells/mm³ or CD4% drops below 25% may not delay the onset of therapy very much, and may risk a child becoming LTFU and/or not having his/her repeat CD4 measured, thus not starting on treatment promptly after dropping below the initiation threshold. However, it would be important to ensure that immediate ART initiation in children aged 2–5 y does not happen at the expense of prioritising early infant diagnosis and rapid initiation, which has been shown to reduce mortality but may be more challenging to implement. This consideration is especially important in resource-limited settings, where an increasing number of patients often overburdens health care facilities, and clinics are therefore not able to trace children who miss their infant diagnostic testing and results or CD4 testing appointments [37],[38].

It has also been suggested that early ART initiation may reduce the high loss to follow-up frequently observed in pre-ART care [39]. However, loss to follow-up has not consistently been shown to be lower in children commenced on ART compared to those not on ART [40]. In addition, there may be selection bias, as children more likely to remain in care are more likely to get started on therapy.

Our results indicate that in children aged 2–5 y in southern Africa, there is no difference in mortality between starting ART immediately and deferring until CD4% drops below 25% or CD4 count drops below 750 cells/mm³. The overall clinical and programmatic implications of earlier ART initiation, such as effects on morbidity-related outcomes, retention in care, cost-effectiveness, adherence, and drug resistance, remain to be explored.