NA ACCORD results support the importance of the START study: response to researchers

Nathan Geffen, TAC

The following article was produced by the lead investigators in the START trial but has wider relevance for clinicians and patients following discussion about when to start treatment and the particular importance for data based on randomised studies.

Note to investigators from Andrew Phillips (START statistician), James Neaton (INSIGHT principal investigator) and Abdel Babiker, Sean Emery, Fred Gordin & Jens Lundgren (START protocol co-chairs)

A recently published analysis of a large multi-cohort dataset (NA ACCORD) has considered the question of earlier initiation of ART. [1] Using two starting strata in separate analyses, CD4 350-500/mm3 and > 500/mm3, the investigators reported lower death rates in patients initiating therapy early compared with those deferring.

The NA ACCORD findings appear consistent with the evidence from epidemiologic studies that were the motivation for the design of START. We did not consider that those data were sufficiently compelling to recommend any change in guidelines to initiation of ART at CD4 count above 350/mm3 in the absence of a trial. That remains our view with the addition of the NA ACCORD analyses to the body of evidence.

On the contrary, we think that the findings from NA ACCORD provide further interesting epidemiological analyses that support the need to perform a randomised trial to assess whether ART should be initiated at higher levels than is currently the case. Specific comments on the NA ACCORD analyses are given below.

The primary reason why findings from large randomised controlled trials are considered the most reliable form of evidence for the impact of an intervention is that unknown and unmeasured confounders would be balanced between arms. No analyses from observational studies can ensure this balance.

It would be a dangerous precedent for medical research if we were to allow our interpretation of analyses from observational data to undermine our ability to perform randomised studies of critical questions of enduring and world-wide importance.

Specific comments on the NA ACCORD analyses

1. As stated by Sax et al in their commentary on the NA ACCORD article, unmeasured confounding could well be strong. People who start ART early are likely to be different to those who do not, particularly in terms of socio-economic status, adherence and health-seeking behavior in general. Guidelines state that the perceived likelihood that a patient will adhere should be taken into account when deciding whether to start ART. [2]

Many aspects of health seeking behavior cannot readily be measured but are likely to be strongly linked to mortality and thus will confound the association between early therapy and mortality. Adherence to placebo has been found to be strongly associated with reduced mortality in several studies. [3-6]

In addition, since the study was not specifically designed to address the question, factors known to predict death from non-AIDS conditions, such as smoking, diabetes, hypertension, and HBV status were not measured and adjusted for, while most deaths of those with cause known were from non-AIDS conditions. Such confounders could jointly well be strong enough to remove the association observed; i.e. with a joint effect which is stronger than the effect of the single unmeasured confounder considered in the NA ACCORD analysis.

2. There are other potential biases in the analysis. Patients in the early therapy group are more likely to be those with a lower natural rate of CD4 count decline and hence better prognosis. For example, consider the analysis of people with baseline CD4 count > 500, and further consider a patient within this analysis who starts ART on the date of the next CD4 count after baseline. If this next CD4 count is > 500 then they will become a patient in the early-therapy group while if the next CD4 count is < 500 they will become a patient in the deferred-therapy group. Thus patients who have a more rapid decline in CD4 count between baseline and the next count are more likely to get into the deferred group and patients with a stable CD4 count are more likely to get into the immediate group. Once the next CD4 count after baseline has been performed, the only way of a patient being in the early-therapy group is if they have demonstrated relative stability in natural CD4 count decline, a factor associated with better prognosis. The extent of the bias caused, and whether it is accounted for by further inverse probability weighting, is unclear.

Another concern is that baseline is later in calendar time in the early-therapy group than in the deferred group. The calendar dates at baseline are not given but the year of starting ART is similar in those in the early therapy and deferred therapy groups in both analyses, suggesting that baseline is on average at an earlier date in the deferred therapy group compared with the early therapy group. This could lead to a bias, which would be removed by adjustment for calendar date of baseline in the analysis. However, simultaneous adjustment for both dates would not be possible.

3. The analysis is opaque, even to statisticians. The methods used are such that it is not possible to show any descriptive data that intuitively give a feel for the higher risk in the deferred group. Based on the number of deaths and person years given the crude death rates are lower for the deferred-therapy group than in the early therapy group in both analyses. The stratification by cohort or use of inverse probability of censoring weighting leads to the relative rates reversing in direction, for reasons that are not easy to intuitively understand. The weighting is used to adjust for censoring due to people starting ART or follow-up ending while still above the CD4 threshold but beyond 6 months from baseline, and for people having a CD4 count that declines below the lower threshold for deferral. Factors associated with loss to follow-up may be different in each case and it is not clear if the same weighting is used. Likewise, the artificial censoring means that there is 2-3 times longer median follow-up in the immediate therapy group, the implications of which are unclear.

4. The analysis in the lower strata is not relevant to START. In the analysis of people with baseline CD4 count 350-500/mm3, the CD4 count at start of ART in the deferred group is below 228/mm3 in 25% of patients starting ART in this arm. This means that this analysis is not relevant to whether ART initiation should be at CD4 counts above 350/mm3, compared with 350/mm3 itself.

5. Another analysis of large collaborative cohorts of the question of early ART initiation carried out recently, using different methods pioneered within the MACS cohort, has concluded that a CD4 count of 350/mm3 is the minimum threshold at which ART should be started. There was no evidence for a mortality benefit from starting at higher CD4 count levels. The reasons for the difference in findings are unclear.


  1. Kitahata et al. Effect of Early versus Deferred Antiretroviral Therapy for HIV on Survival. Early release at April 1, 2009 (10.1056/NEJMoa0807252).
  2. Panel on Antiretroviral Guidelines for Adults and Adolescents. Guidelines for the use of antiretroviral agents in HIV-1-infected adults and adolescents. Department of Health and Human Services. November 3, 2008; 1-139. Available at
    Accessed 6 April 2009.
  3. Canner PL. Influence of adherence to treatment and response of cholesterol on mortality in the Coronary Drug Project. N Engl J Med 1980; 1038.
  4. Irvine J, Baker B, Smith J, et al. Poor adherence to placebo or amiodarone therapy predicts mortality: Results from the CAMIAT study. Psychometric Med 1999; 61:566-575.
  5. Simpson, SH; Eurich, DT; Majumdar, SR, et al. A meta-analysis of the association between adherence to drug therapy and mortality. BMJ 2006; 333:15-18.
  6. Granger, BB; Swedberg, K; Ekman, I, et al. Adherence to candesartan and placebo and outcomes in chronic heart failure in the CHARM programme: double-blind, randomised, controlled clinical trial. Lancet 2005; 366:2005-2011.

Links to other websites are current at date of posting but not maintained.