t*******2 发帖数: 384 | 1 Does anyone know if the overall power will increase or decrease compared
with single stage
analysis? Thanks with Baozi! |
n****t 发帖数: 182 | 2 It depends...
If your interim analysis is futility analysis, by definition, you may
wrongly terminate the experiment when your H1 is true. This increases the
type II error, which lowers the power. But if your trial designed correctly,
this can be adjusted (for example, increase sample size). For example, beta
-spending method.
If your interim analysis is efficacy analysis, you'll have more chances to
claim efficacy and this inflates type I error and shrink type II error. Your
power hence increases. This is the multiplicity issue and multiplicity
needs to be adjusted. When type I error is correctly controlled at the same
level as single stage with the same sample size, your power will be lower
comparing to single stage trial. |
t*******2 发帖数: 384 | 3 Thanks, could you elaborate the statement "When type I error is correctly
controlled at the same level as single stage with the same sample size, your
power will be lower comparing to single stage trial." That is exactly what
I am thinking. |
t*******2 发帖数: 384 | 4 Thanks, could you elaborate the statement "When type I error is correctly
controlled at the same level as single stage with the same sample size, your
power will be lower comparing to single stage trial." That is exactly what
I am thinking. |
n****t 发帖数: 182 | 5 For group sequential designs, n.GS/n.fix>1 to remain the same power, where n
.GS is the group sequential designs(with correctly controlled type I error).
For certain designs there are even formula for this.
Google "group sequential designs" should yield a lot of results on this. |
g******n 发帖数: 339 | 6 One thing that you did not mention: for GS approach, there is always a
chance to stop early for efficacy. The sample size you mentioned below for
GS is the maximum sample size, or the expected sample size?
n
).
【在 n****t 的大作中提到】 : For group sequential designs, n.GS/n.fix>1 to remain the same power, where n : .GS is the group sequential designs(with correctly controlled type I error). : For certain designs there are even formula for this. : Google "group sequential designs" should yield a lot of results on this.
|
t*******2 发帖数: 384 | 7 Good point
【在 g******n 的大作中提到】 : One thing that you did not mention: for GS approach, there is always a : chance to stop early for efficacy. The sample size you mentioned below for : GS is the maximum sample size, or the expected sample size? : : n : ).
|
n****t 发帖数: 182 | 8 That's a good point; only the maximum sample size is larger otherwise will
be a bad procedure to use.
【在 g******n 的大作中提到】 : One thing that you did not mention: for GS approach, there is always a : chance to stop early for efficacy. The sample size you mentioned below for : GS is the maximum sample size, or the expected sample size? : : n : ).
|
t*******2 发帖数: 384 | 9 Any result on the expected sample size then? |
n****t 发帖数: 182 | 10 That will depend on which procedure you use: Pocock for example, is
aggressive in interim but O'Brien-Fleming is less so, give you different exp
sample size under different hypothesis.
These probabilities of efficacy can be calculated using software
【在 t*******2 的大作中提到】 : Any result on the expected sample size then?
|