Knut's proposal might not be the most efficient one, but it can be realized. I understand that stopping only for futility is planned and the (only) interim takes place after 80 subjects (20+60) have reached the endpoint.
If H1 is true and thus p2=0.25 then the probability to see an event rate of .45 or higher among the 60 subjects is negligible (p=0.0005925201888). This small probability would hardly influence the power. Things would of course be different if the futility bound decreases.
------------------------------
Joachim Roehmel
------------------------------
Original Message:
Sent: 06-01-2015 02:06
From: Constantine Daskalakis
Subject: power calculation for futility analysis
Knut,
To answer your problem you need to specify
1. Will the interim analysis be only for futility or also for efficacy?
2. Will there be one look or more?
3. When will you be doing the interim analyses? After how many accrued?
Also, why are you choosing the criterion of p2>.45 as futility boundary? That does not seem right.
As far as I know, futility boundaries are computed formally on the test statistic via methods analogous to interim analyses for efficacy (error spending functions). Alternatively, you can do futility assessment based on conditional power.
------------------------------
Constantine Daskalakis
Thomas Jefferson University, Philadelphia, PA
------------------------------
Original Message:
Sent: 05-29-2015 17:14
From: Knut Wittkowski
Subject: power calculation for futility analysis
I'm trying to determine the loss of power from an interim analysis for futility (binary outcome).
I'm planning a study with 200 subjects randomized (1 placebo):(3 verum) and would like to stop early if there isn't at least some benefit in the verum group. H0: p1=p2=.5; H1: p1=2*p2=.5 (no continuity "correction").
|
p1 |
p2 |
d |
alpha |
power |
n1 |
n2 |
|
|
.50 |
.25 |
-.25 |
.05 |
.90 |
50 |
150 |
|
How much more subjects do I need to compensate for loss of power when stopping for futility if p2>=.45 (e.g.) after 20+60=80 subjects? My gut feeling is that 60+180=240 would be more than enough, but I'd like to be a bit more precise.
Thanks
Knut
------------------------------
Knut Wittkowski
Head, Dept. Biostatistics, Epidemiology, and Research Design
The Rockefeller University
------------------------------