I’m investigating a method that’s gaining popularity in epilepsy trials, which is the “time to pre-randomization monthly seizure count” analysis (Reference: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4442101/).
The analysis assesses how long it takes for each patient to reach the number of seizures they typically experienced over 28 days prior to randomization. Rationale: if treatment had no antiseizure effect, patients would remain in a trial for an average of 28 days or reach their pre-randomization seizure count in about 28 days; but if seizure frequency was reduced, patients would continue treatment for longer and it would take longer than an average of 28 days to reach their pre-randomization seizure count.
The time-to-event (TTE) endpoint is defined by the number of days until the subject cumulatively reaches/exceeds their pre-randomization monthly seizure count.
- The response variable is coded to 1 (yes for reaching the TTE endpoint) and 0 (no/censored).
- Subjects who discontinue the study before reaching the endpoint are censored at the time of withdrawal.
- Kaplan-Meier (KM) analyses are used to estimate the median time to pre-randomization monthly seizure count and 95% CIs.
- The log-rank test is used to assess the significance of any difference in time to pre-randomization monthly seizure count between the active group and placebo.
My thoughts
- Biased results with increasing dropouts by coding their response variable to “censored”, as their cumulative seizure counts after withdrawal may very well reach/exceed what they typically experienced over the pre-randomization period but cannot be accounted for in this analysis, not to mention the reasons of withdrawals are likely informative. Wouldn’t it be better or more conservative to code the response to 1 (ie, the bad outcome) in this case?
- Instead of KM and the log-rank test, perhaps the authors could have used Cox regression. Their current analysis does not consider baseline and other covariates.
- There is great within-subject variability in the number of seizures experienced by the subjects. A baseline period of 28 days may not sufficiently characterize what is typically experienced by someone pre-randomization. It would be better (though challenging) to obtain data over several baseline periods or from a longer baseline period pre-randomization. The authored mentioned that sudden unexplained death is an issue for patients randomized to placebo so it may no longer be reasonable to expect patients to accept the possibility of randomization to placebo for 3-4 months.
o Since death is mentioned, competing risk(s) should be considered in the TTE analysis, which cannot be handled by KM.
I would appreciate any feedback (pros, cons, and improvements) regarding this TTE approach. Many thanks.