Study Design
Strategies to Avoid Immortal Time when Emulating Sequential Trial Designs Using Observational Data Michael Webster-Clark* Michael Webster-Clark Webster-Clark Webster-Clark Webster-Clark Wake Forest University School of Medicine
Emulating hypothetical randomized trials is an increasingly popular framework for non-experimental studies. In longitudinal data, emulating a series (or sequence) of randomized trials and pooling results may be more efficient than emulating one randomized trial. The prevalent new-user design, for example, emulates a series of trials where individuals initiate a new treatment strategy or continue on an old treatment (comparator). When trials are separated based on time intervals (e.g., switch in month 1, switch in month 2, switch in month 3), immortal time can be introduced if individuals cannot be assigned to one group prior to the end of the interval.
We propose three solutions with unique implications for the target trial under study (Figure 1): A) start follow-up on index events that occur within intervals (e.g., dates of prescriptions for the treatment of interest or comparator in a prevalent new-user study); B) assign treatment group among interval survivors and start follow-up at the end of the interval, or C) clone, censor, and weight each hypothetical trial. We also evaluated a solution D) that started follow-up at the beginning of the interval and looked into the future to assign exposure group. In a simulation study with a null treatment effect, solutions A, B, and C estimated a null 1-year risk difference comparing the two treatments; D did not.
Each approach has limitations. Clear index events (A) may not always exist (e.g., when the “comparator” involves no intervention). Starting follow-up at the end of intervals (B) separates treatment assignment, exposure, and the start of follow-up. Cloning and censoring (C) estimates “naturalistic” effects conditional on the population and pooling trials further complicates interpretation. Both (B) and (C) become more problematic as time interval become longer. Regardless, researchers should try to avoid introducing immortal time when analyzing observational data as a series or sequence of trials.
