We have a new whole-brain fNIRS system, where we can analyze 108 channels, total. We often do explanatory research having no a priori ROIs for our fNIRS data. Interested in channel-wise analysis, we end up having 108 tests we theoretically would have to correct for. Even with FDR-correction, we need at least one p-value
If you have a rough idea of effect size, you could try a power analysis. You might need many more participants than you are currently using. Your p-values are influenced by degrees of freedom imposed by your sample size for the type of univariate approach you are describing.
This problem arises in the analysis of DNA microarrays (eg https://www.amazon.com/Cancer-Diagnostics-Microarrays-Steen-Knudsen/dp/0471784079/ref=sr_1_5?keywords=Steen+Knudsen&qid=1579659449&s=books&sr=1-5 ) One solution used there is given by the Benjamini-Hochberg procedure instead of Bonferroni. Please see the attached. Best wishes, David Booth
This is a known issue when dealing with multi-channel or multi-voxel brain imaging. The main problem arises from the statistical dependance of the results as a function of space, namely, close channels or voxels are highly correlated. This can drastically inflate your false negatives when using standard procedures for multiple comparison correction. In my opinion the best approach is to actually account for spatial statistical dependence using, for example, gaussian random field theory. The approach will provide an estimation of the 'effective' number of independent measures. This is the standard procedure in fMRI where there are a huge amounts of voxels to deal with. If i remember correctly, nirs_spm had an implementation of this approach, however the software is getting outdated.
I agree with Antonio that this is similar to the multi-voxel comparison in fMRI. Therefore, most solutions addressing that problem, could be used for your problem with caution, depending on the exact question you are trying to answer (i.e. there is no best solution for this problem in my view).
Week 4 of this course on Coursera gives some insight on Multi-voxel Pattern Analysis (MVPA):
https://www.coursera.org/learn/functional-mri-2/home/week/4
Hi everybody,
thank you very much for your various suggestions. I am very thankful for every advice helping us to come up with a proper solution.
I will try to add all mentioned approaches in my answer, however, if I forget to mention some, please let me know. Also please feel free to recomment if I didn´t get you right or if you do not share my opinion on certain things.
First, we usually add a power analysis before starting data acquisition, which helps selecting our sample sizes according to our expected effects. However, sometimes we are doing whole-brain fNIRS recordings while subjects perform sports-related movements which have not been reported, previously. Therefore, previous studies are missing and we are rather doing explanatory research. In these cases, I suspect that a priori sample size calculations are impossible?!
Maybe I did not mention it giving enough detail but we are already trying FDR (Benjamini-Hochberg) correction since this approach is more liberal compared to Bonferoni. However, the most significant finding must still pass the bonferoni correction value - and this is what makes statistical testing so complicated because this value is really small having this large channel number.
As far as I understood, gaussian field theory and the associated spm approach relies on normally distributed data, is this correct? Unfortunately, fNIRS data often is NOT normally distributed and therefore the results may be biased by the non-normality of the data. I don´t know nirs_spm too wel so there might have been a solution in the past years, I did not realize. Are you aware of any corrections for non-normality in nirs_spm?
Multi-voxel pattern analysis is also an interesting approach, I did not think of, yet. Thanks a lot for sharing the link, I will have a look and see what we may use for NIRS data analysis from it.
What do you guys think of an interesting idea, a colleague of mine shared: namely calculating a multiple regression analysis having the channels as predictors of the behaviour, we are measuring. Having the betas, we can define thresholds, which amount of explained variance should be minimally explained and subsequently only add these channels in a permutation analysis?
Thanks everyone again for participating in this discussion! I am very honoured, you share your experiences and advices with me. Any follow-up is highly appreciated!
Elisabeth
Hi Elisabeth,
If you want to continue with FDR, you can try the Storey (2011) approach which is more liberal than Benjamini-Hochberg. The R package and Shiny are very user-friendly: https://github.com/StoreyLab/qvalue. You might want to consider creating another question regarding the permutation analysis. Hope you find a solid approach!
Best,
Justin
- Badges
- Science topic
- Similar topics
- Correction