Can I acquire p-value thresholded nii from 'The Decoding Toolbox' output?

Hi, I’m currently trying to use ‘The Decoding Toolbox’ (TDT).

Since our fmri experiment conducted only 2 runs for each subject,
I used decoding_between_subject template from TDT.

I set cfg (configuration options) as searchlight analysis, and accuracy_minus_chance output.

Next on, I tried to plot the nii file with accuracy_minus_chance score,
but it doesn’t seem to survive from p-value 0.03 (which is an intriguing value, that my data is comprised of 30 participants).

So the question is:

  1. Can I modify my codes within TDT and get the appropriate accuracy scored image,
  2. Or is this kind of accuracy map should be plotted within different toolbox offering an adequate option (other than spm plotting)?

p.s.) When I saw some articles using MVPA, guess nii was thresholded with the same criteria (p-value, fwe, …)?

Hi,

I’m really not sure I understand the question, sorry. When you mention the accuracy_minus_chance result, are you saying that the lowest accuracy is 0.03?

I think instead of classifying between subjects, I would recommend the correlation_classifier first and use signed_decision_values as output (decision values here are the z-transformed correlation coefficient of class A minus the z-transformed correlation coefficient of class B). Or you could try crossnobis (check the templates folder).

Here are the relevant settings for the correlation classifier and signed decision values:
cfg.decoding.method = 'classification';
cfg.decoding.software = 'correlation_classifier';
cfg.results.output = {'signed_decision_values'};

Best,
Martin

Thanks for the fast reply! I really appreciate.

0.03 was the p-value threshold for the res_accuracy_minus_chance.nii file when I tried to plot it with spm gui, not the accuracy.
Actually the accuracy score from each voxel seemed proper… (just gut feeling though TT)
What I meant suspicious was that the reason there didn’t exist significant voxels below p-value 0.03 might due to the 30 subjects (0.03 ~= 1/30).

Anyway, as you mentioned, there isn’t a way to transform this accuracy_scored_map into z-scored_map?

One more last question… Can’t I choose cfg.decoding.method = 'searchlight' along with 'signed_decision_value'???

I’m not sure this makes a lot of sense really, unless you also ran a statistical analysis somewhere to get p-maps (e.g. using permutation testing).

Unfortunately, you cannot treat accuracies as statistical values. You first need to run a statistical test to see if your accuracies are significantly above chance. One approach is to run a permutation test.

I think it would be best if you just tried it out :wink: (but short, yes, it should work).

Best,
Martin

1 Like

Thanks soooo much!!! :sweat_smile: :slightly_smiling_face: :star_struck: