PURPOSE: To describe a bias that can occur in the analysis of data from certain randomized trials.
METHODS AND RESULTS: Although randomized trials are effective at preventing confounding, a potentially strong confounding can arise in certain therapeutic drug trials in which follow-up is extended in an open-label phase that allows switching treatments. Many trials implement eligibility screening that excludes those at high risk of morbidity and mortality. In these trials, disease rates and death rates for the stud population can rise rapidly during follow-up as the effect of screening wanes. During the open-label follow-up, a preponderance of patients may switch to the new therapy. If so, then any evaluation of the new therapy that includes follow-up from the open-label phase, as is often the case for safety evaluations, will be confounded. The confounding arises because the person-time experience of those on the new treatment will be more heavily weighted with the open-label phase experience, during which morbidity and mortality rates may be much greater than in the initial phase of follow-up. This confounding may be strong and will be in the direction of making the new treatment look worse, provided that the net switching is toward the new treatment during the open-label phase.
CONCLUSIONS: The confounding described here is not prevented by randomization because it develops in a non-randomized add-on analysis to the trial. The bias can be removed, however, by controlling for time since randomization in the analysis of the data.
