Mensen
commented
6 years ago Hey!
1) Difficult question. Power analyses are essentially working backwards from a traditional statistical test and getting to the minimum N. However with non-parametric tests (which at the statistical level the TFCE approach is), is more difficult and many more assumptions need to be made since these sorts of approaches use all the data points (and not just the mean and standard deviation like parametric tests do).
For example, see this link, to see the sort of discussion and nuance necessary when dealing with power calculations for the more straightforward test.
Of course the TFCE approach complicates this sort of picture even further since you would also have to make assumptions about the level of neighbourhood support you expect to see in your data... which I'd be super skeptical you could make with any degree of certainty.
So ultimately, you guesses as to what your data will look like will have a huge error in them. Any upper and lower bounds you might set on these with a reasonable degree of certainty would probably leave you with N estimates from say "you need either 5 participants... or 400 for this approach to be useful"... and so would be useless.
2) You always have the underlying values so you can still calculate cohen's d... and this would be accurate for what its worth. But in doing so you wouldn't be computing the amount of support for that particular channel/sample which is one of the motivations behind TFCE in the first place. But short of reinventing the wheel on this topic... I would hope that reporting the actual mean difference, and in these cases, the number of significant channels and/or time points (or frequency bins)... will give the reader all the necessary information to make a good decision on whether your effect is not only statistically interesting, but the actual difference is also of interest.
Not great answers I now... but I don't have a lot of faith in power measurements... nor in standardised effect sizes because usually the most important factor in your actual difference is dependent on other studies in the field and the consequences of that amount of difference. E.g. if I tell you slow wave amplitude increased by 15 uV for condition B (and a small cohen's d) you might not know what to make of it. If, on the other hand, you know that total sleep deprivation has a 10 uV increase then you'll know that 15 uV is a lot and can make a better judgement.
81N55E
Hey Armand, sorry - me again. Two questions: (1) A-priori power analysis: Just as an example... in G*Power for a fixed effect with interactions ANOVA with an expected effect size of .25, alpha of .05 and Power of .80 and 2 Groups I have a suggested total sample size of 269. Would the same sample size stand with the TFCE approach? If not, how can I do a a priori power analysis for the TFCE? (2) Effect size: Is there a possibility to implement in the results-output the Effect size and CI as it tells me more about the actual effect than the p-value? Best & Thanks!, Mitja