Re: Errors Regarding Interpretation of Trial Stopping Early
Hi, statistician here, thank you for bringing up trial early stop issue, but there are a few errors in your analysis.
1. O'Brien-Fleming stopping criteria has a maximum number of final + interim analysis (5). You can easily see this if you look up the OBF table, as the p recommended for stopping the trial after the 5th analysis is around 0.04.
2. Since 2013, IDMC has twice a year for interim reports, so until now, there are probably more than 14 analysis. This is very important, as the probability of making a type I error via interim analysis is approximately 5% with 5 views, but increases to nearly 20% with ~15 views, this means both the p value and hazard ratios must be adjusted much, much lower than the values you quoted (which didn't align with OBF).
3. Re: Fosco's Model, if there were 162 event in control by Sept 2019, there are probably 109-126 events in the SOC as per his model, dividing the two number (assuming 1:1 randomization ratio) yields a HR in 0.60s-0.70s range, idk how you were able to derive 0.84.
4. My own Kaplan Meier survival shows the HR to be ~0.7 with p<0.0001 via a chi sq. log rank p test, again, idk where 0.84 comes from, please explain.
5. (Perhaps) most importantly, the trial assumes 85-90% stat p, while OBF assumes 90% stat p for the endpoints for stopping. Hence, if the dropout rate was greater than 10%, then the stop boundaries would need to decrease again for the trial to be stopped early for efficacy. Hence, an IDMC would still recommend a trial to continue even if 4. was met.
TLDR, great job discovering OBF, but there are a few problems with your statistical interpretation of it.
Looking forward to your reply!
Market Data 
Markets 
AI
