[Repost] All voxels in the whole brain show above chance level in TDT result

1

Hello TDT users,

I uploaded this post in the past, but there was no reply so I re-posted it.

I have 16 beta images for each of the 3 conditions across four runs per subject, which were used as the input for individual’s searchlight analysis.
I predicted that some decodable regions will show positive value in the file ‘res_accuracy_minus_chance.nii’, but all voxels in the whole brain show above chance level in TDT results unexpectedly. (I uploaded the file as example.)
So, I wonder if this result is valid, or if I set wrong parameters for decoding analysis.

The parameters that I used is as below:

cfg = decoding_defaults(cfg);
cfg.testmode = 0;
cfg.analysis = decoding_type;

cfg.decoding.method = ‘classification_kernel’;
cfg.decoding.software = ‘libsvm’;
cfg.decoding.train.classification.model_parameters = ‘-s 0 -t 0 -c 1 -b 0 -q’;

cfg.software = spm(‘ver’);

cfg.results.dir = output_dir;

cfg.searchlight.radius = 12;
cfg.searchlight.unit = ‘mm’;

cfg.files.mask = fullfile(mask_dir);

cfg.files.name = alldesign(:,1);
cfg.files.label = cat(1, alldesign{:,2});
cfg.files.chunk = cat(1, alldesign{:,3});
%(I checked that the beta images are allocated well in the design matrix.)

cfg.design = make_design_cv(cfg);
cfg.results.output = {‘accuracy’,‘accuracy_minus_chance’};

results = decoding(cfg);

Any feedback would be great, Thanks!

~ Taehyun Yoo

res_accuracy_minus_chance(red)_on_T1_image
res_accuracy_minus_chance(red)_on_T1_image1851×755 184 KB

2

Hi Taehyun,

Apologies for the delay in the response. If you checked that the betas are definitely correct, then it is possible for these things to happen when there is a strong difference in the estimability of beta coefficients from regressors for one condition compared to another. Beta coefficients will be more variable when there is a higher covariance between some regressors than other or when they have less energy (i.e. “there is less going on”) in some regressors than others, e.g. because one condition has many fewer trials than another. So, possibly, condition A has many more trials than condition B. We highlight this effect in this paper in Fig. 5, and provide a more general explanation in this paper in Fig. 7.

Unfortunately, there is not an easy solution to this problem that works in general and for everything, but the approach suggested here might be feasible.

Best,
Martin

3

Dear Martin,

Thanks for your fast and kind answer!
The literature which you cited was very helpful!!
Still, I am struggling with this topic, but I will try to figure it out.
I will ask whenever I have some questions.

Thank you!

~Taehyun Yoo