Accuracy and precision of estimating intervention efficacy when the timing of observed events differ by treatment arm.

Randomization in clinical trials minimizes differences between treatment arms, allowing observed treatment differences to be attributable to an intervention. For prospective clinical trials, we examine the effects on inference when other specified treatment arm differences are also present. These differences are imposed using three measures: time between the unobserved failure event (e.g., HIV infection) and its detection, visit schedule adherence and dropout. Our context of interest is trials with non-recurrent time-to-event outcomes and fixed visit intervals, where treatment efficacy is measured either by a hazard ratio or by a ratio of cumulative incidence functions. Moderate treatment differences in visit adherence, either through missed visits or additional unscheduled visits, were not found to cause substantial bias or to reduce power. However, both differential loss to follow-up (when coincidentally dependent on risk of failure) and differential time between event and detection should be of concern in designing clinical trials. Efforts to re-capture subjects at the end of study for failure assessment are helpful in some contexts, and may be considered in study planning.