Dear Chris and Joke,
Ok thanks. Although the suggested procedures mean that I need to recruit more subjects, It really sounds better and more reasonable! I have no idea if everyone doing fmri study does do it this way (I.e., get first group, calculate power, then get another group). I guess most people just collect "enough" data and then publish it. What "enough" means is just statistically significant n number.
Mike
從我的 iPhone 傳送
> Joke Durnez <[log in to unmask]> 於 2015年8月23日 上午8:42 寫道:
>
> As Chris says, what you describe is referred to as the unsound practice of data peeking.
> This procedure will inflate your type I error rate. So while you correct in each step for FWE and your significance level is at alpha=0.05 every time, the final familywise error rate of your analysis is a lot higher than 0.05!
>
> My suggestion:
> (1) use only the 10 first subjects
> (2) perform a group analyses
> (3) upload the resulting statistical map to https://neuropower.shinyapps.io/neuropower, and follow the instructions. Now you’ll now how many subjects you’ll need for a high-powered study
> (4) Discard these 10 subjects and use only the subsequent subjects for your final analysis !! This seems hard, but it is important that the sample on which you based the power analyses are independent from the final sample, otherwise this will again inflate your type I error rate.
>
> Joke
|