Replies to post #158659 on NorthWest Biotherapeutics Inc (NWBO)

[AVII77]

AVII, so how many months difference between the cohorts would you need for stat sig of .05 or less with the event size as we now have it?

Please see my post here where I calculated that:

To hit stat sig here on OS (after the "enhancements") with a p of 0.05 after observing 233 OS events requires an improvement of 5.3 months (beyond 17 in control).

Also note that "5.3 month (beyond 17 in control)" is not the same as "5.3 month (beyond 22 in control)".

In the first case (assuming proportional hazards) the HR is 0.76 while in the second case it is .80.

The smaller the difference, the more events it takes to reliably conclude the difference is not observed by chance.

That's why trials powered to detect modest, but important, differences (say 15% RRR in AMRN's R-IT trial) require many (1600 for R-IT) events. (and R-IT is powered at 90%, not 80%. That too increases the event count needed).

Bottom line, with 233 events, the observed difference between arms must be substantial (about a 24% reduction in the risk of death).

And yes, my calculations assume proportional hazards (no long tail) because the math is easier. But if there is a long tail (non-proportional hazards) that means events occur on both arms prior to the curves diverging. Those events don't "help" differentiating any treatment benefit. So my numbers above are for "discussion" and are conservative; a larger difference may well be needed if there is a "long tail" - and I don't mean an extra 20 events.

As Hoos notes:
“As a consequence, computation of the required number of events for final analysis under proportional hazards assumptions will lead to a statistical power insufficient for a trial with a delayed separation. Depending on the timing of the delay, this loss of statistical power can be substantial.”
http://jnci.oxfordjournals.org/content/early/2010/09/08/jnci.djq310.full.pdf