Loss to follow-up (LTFU) in clinical trials represents a potential threat to their soundness that may not be adequately recognized. We consider a log-rank test in a trial with two arms, experimental and control, and with a single unfavorable binary endpoint such as death. Commonly, one applies censoring to patients with LTFU. That approach is valid if LTFU is independent of outcome, but can lead to bias otherwise. Unfortunately, there is no statistical test for independence, so the legitimacy of the approach rests on unverifiable assumptions. For two cases, we evaluate the impact of the approach based on simulations that use reasonable models for outcome-dependent LTFU. In each case, LTFU in one arm disproportionately suppresses recognition of relatively early deaths or other outcomes, thus producing bias favoring that arm. The first case has extra LTFU in the experimental arm and the treatment has no benefit. The second case has extra LTFU in the control arm and the treatment is effective. The simulation results show severe inflation of Type I error in the first case and major loss of power in the second case. Remedies for LTFU are scarce but include avoiding it in the first place where possible.
Keywords: Bias; clinical trials; log-rank test; loss to follow-up; simulations.