Decoding for continuous variables

Hi TDT users,

I noticed that most multivariate decoding studies used a binary decoding approach. The dependent variables were continuous in some studies, but they still did a binary transformation via median split. I am curious if such multivariate binary decoding has any advantages in general?

Also, I was wondering if it is possible to use the TDT to decode continuous variable (similar to a parametric analysis for the univariate approach)? Is there a template for this?

Thanks!
Yuanwei

Hi,
My response is not related to TDT, but in geenral, you can perfoectly wtudy high-dimeniosnal regression using tools such as Risge Regeression or Support Vector Regression. This is actually computationally simpler than classification !
The point is that sometimes the actual values are not very accurate or meaningful, and thus in many cases, it is a better idea to frame the problem as classification than regression.
HTH,
Bertrand

Hi Bertrand,

Thanks for your reply! I also noticed that some studies claimed that multivariate binary decoding may yield more reliable results than regression in some cases. Are there any MVPA studies that explicitly addressed this issue and specified in which situation we should do a binary classification rather than regression?

Another related question: I used to do a categorical analysis (e.g., high vs. low value) at the first level before doing multivariate binary decoding, but I’m not sure how to control for the effects of trial-wise confounding variables in this case. Is it possible to do a parametric analysis and include both the variable of interest (e.g., 1 and -1 for high and low value) and confounding variables (e.g., response time for each trial) as parametric modulators at the first level, and use the beta weights of the variable of interest for the subsequent decoding?

Thanks!
Yuanwei

Sorry, I’m not aware of such papers.
Yes, deconfounding your analysis at the first-level analysis step is a good idea.
Another possibility is to craft test sets on which RT is independent from difficulty (see Chyzyk et al. GigaScience 2022).
HTH,
Bertrand

Hi Bertrand,

I have a question related to deconfounding: can I add the confound (RT) as nuisance regressors (same as head motion parameters). If so, should I add RT values to only the event onset or its whole duration?

Thanks!

Frodo