Janet Darbyshire and colleagues from the MRC Clinical Trials Unit in London pioneered the conduct of a series of ground-breaking HIV treatment trials during the 1990s. One of these trials, The Delta trial randomised more than 3,200 patients to treatment with zidovudine (ZDV) alone or ZDV in combination with either didanosine (DDI) or zalcitabine (DDC). One of the questions the trial wanted to answer was whether combination treatment with DDI was superior to DDC.
The authors analysed this in two different ways. The first approach looked at the direct randomised comparison of patients randomised to either DDI or DDC. The second method identified those patients initially randomised to just ZDV and whose treatment was subsequently modified in an open-label fashion to include combination treatment with either DDI or DDC. Death was chosen as the study endpoint with the groups being similar in terms of selected risk factors for clinical progression. This means that the groups “looked” the same giving a false reassurance to be able to reliably investigate the effects of combination treatment across both the randomised and observational analysis.
The analysis clearly showed that in the randomised group ZDV plus DDI showed improved survival compared to ZDV plus DDC, but in the observation group the opposite effect was seen, namely combination treatment with DDC showed improved survival over DDI. Why might this be the case?
One explanation suggested by the authors is that in the observational group patients with a poorer prognosis are more likely to be prescribed DDI over DDC. This selection bias was not identified from the baseline analysis of selected risk factors for progression highlighting again the false reassurance that such results can confer. This highlights once again that only with a proper randomised comparison can selection bias resulting from known and unknown characteristics be reliably eliminated.
You can read the full paper here.