ADVANCE and glycaemia thresholds: a need to clarify the statistical approach

To the Editor: The recent study published by the ADVANCE Collaborative Group confirming that there is a lower threshold of optimal HbA_1c [1] is a welcome confirmation of the original work published in The Lancet [2]. However, there is at least one important analytical failing that simply must be addressed for this study to be considered valid. Furthermore, the authors’ analysis may violate an assumption of the survival methods that were applied in the original study.

The impact of HbA_1c on vascular outcome can only have one true association. However, Zoungas and colleagues report two consistently differing patterns of association dependent upon whether participants were randomised to their intensive treatment arm or to standard care (see Fig. 1 in Zoungas et al [1]). Explained simply, the association between the hazard of major macrovascular events and HbA_1c was higher at lower HbA_1c levels in the intensive treatment arm. These systematic differences must relate to residual confounding. The authors are aware of this potential problem since they point out in the Discussion section that ‘…despite extensive multiple variable adjustment, these analyses may be unable to completely eliminate the effects of residual confounding attributable to disease severity’.

The original report by the ADVANCE Group published in The New England Journal of Medicine [3] reported clearly that the most striking differences between the treatment arms by the end of the study related to differences in treatment regimens (Table 1). In fact, there was remarkably little difference between the two arms in any of the other reported variables at the end of the study (see Table 1 in ADVANCE Collaborative Group et al [3]). Two of the most obvious differences between the two arms were the increased use of sulfonylureas and/or insulin in the intensive arm (Table 1). Both of these treatments are hypothesised to increase the risk of adverse outcome vs alternative treatments [4, 5].

Table 1 Glucose-lowering treatment regimens of participants enrolled in the ADVANCE study at baseline and at end of study

Regarding the issue of violation of the assumptions of the Cox model [6], HbA_1c appears to have been introduced into the model as the mean of all HbA_1c observations between baseline and the end of the study. This is technically incorrect in that the survival model should not include fixed values following baseline. The original Lancet paper used sensitivity analysis to explore this matter [2]. Whilst I agree that this is unlikely to have impacted on their findings, it is important that HbA_1c be introduced as a time-dependent covariate to check the sensitivity of this covariate. After all, this is the central focus of their epidemiological study.

It is vital that the authors re-run their analysis and evaluate potential treatment-related effects. The fact that this has not been done already is surprising and deserves an explanation, since it is important in terms of the interpretation of both the original study and the secondary analysis. Their discussion point relating to residual confounding in terms of disease severity is bewildering given that the most obvious source of residual confounding is well-documented in their reports.