I'm performing statistical analysis on a time series data, comparing the average amplitude (y axis) between three groups on every 50 ms (x axis). The time series is 1200 ms long (so in total 24 time windows). As I'm repeating the test 24 times, I have to correct the p value for multiple testing. Because the time windows have diferent distributions, sometimes I have to use a normal test (paired Anova) and sometimes I have to apply a non-parametric test (Friedman). At the end I have a list of 24 p values, which have to be corrected for false discovery. Am I allowed to perform a Benjamini Hochberg FDR correction, even though my p values are originated from different tests?
I would appreciate any contribution
Hi Jochen Wilhelm , thanks for your answer. I am indeed considering to treat time as a continuous variable. However I am working with Event relatad potentials from EEGs. A common approach for analysing EEG data is to previously determine a time window and then calculate the average mean of it. I'm doing an exploratory analysis of my data though, and I don't want to pre-specify a time window by now. Therefore, I would like my script to show me whenever there is a significant difference between groups and the aproximate location of these difference would also be an important piece of information to me (average amplitudes pointed out by the time window). If I treat time as a continuous variable I guess I would still have the same multiple comparison problem, because the number of tests would be even bigger (I would be comparing groups at each time point). But maybe I could perform a cluster analysis with family-wise error correction. Would that be your suggestion?
You don't need tests if you do an explorative analysis.
It would be helpful to know what kind of differences you are kooking for. For instance, it is a particular peak that can occur earlier or later, and/or be wider or narrower? Or is is a wave with a particular amplitude and/or frequency that can be distinctively different between groups?
A more general option to find time windows with differences between groups is to fit splines. You can get confidence bands, and time regions where confidence bands don't overlap are good candidate time regions where the profiles might by different.
https://www.unfoldtoolbox.org/contenttutorials.html
https://fromthebottomoftheheap.net/2011/07/21/smoothing-temporally-correlated-data/
https://sites.lsa.umich.edu/karenen/wp-content/uploads/sites/222/2015/03/mssiss2015a0.pdf
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA