Hi TDT team,

I have four beta.img (2 conditions) per subject across two runs that were used as the input for individual’s searchlight analysis. The res_accuracy_minus_chance.mat file generates discrete values (0, -25, 50, etc.), but when I open res_accuracy_minus_chance.nii file in SPM, the values (I’m looking at the intensity value) seem more continuous and do not match with the .mat output. So my questions are:

  1. how does TDT calculate the classification accuracy? Are the discrete values related to how many input files are included (so since I have 4 files, the accuracy ranges from 1/4 to 4/4)?
  2. What’s the relationship between res_accuracy_minus_chance.mat, and res_accuracy_minus_chance.nii? If I feed individual’s accuracy map into SPM for a 2nd-level analysis, are discrete values from the .mat file used as input?
  3. A side question: the second group analysis in SPM generates a beta_0001.img file, and I wonder if this can be used to interpret the result, rather than specifying a contrast vector (i.e., 1 if there’s one group) and using the spmT.img?

Thank you!


Hi RL,

If you only have two runs, I’m not sure this is the best approach. I think I would use the 'correlation_classifier' instead of 'libsvm' and use 'signed_decision_values' instead of 'accuracy_minus_chance'. This will be equivalent to a Haxby-style correlation-based classification analysis and will provide the difference in Fisher z-transformed correlation coefficient between patterns of the same condition and patterns of different conditions.

My guess is that in SPM you were using a view that smooths the data. Right click on that view and possibly turn off smoothing.

It takes all individual predictions and compares them to the individual true labels. Note that this can lead to different results than first calculating the fold-wise accuracy and then averaging those. But in practice, usually the two are the same.

res_accuracy_minus_chance.mat contains all information that is written to the image in res_accuracy_minus_chance.nii. In that respect, it’s identical, but just has more information. It’s mostly for people who want to work on the numbers directly, rather than the images.

I think this is more of an SPM question. A con-image is just the product of your contrast vector and your betas. That means if your contrast is [1], then the con-image is 1*beta, which is identical to the beta.

Hope that helps,