Minimum number of trials

[A. G.]

Hi Cosmoteam,

first of all, thanks a lot for such a nice tool!

I am new to the cosmo universe,and I was wondering the minimum number of trials I need to obtain reliable results. I run a memory paradigm with 4 conditions and 4 runs (so 16 bins). However, in some bins I only have 3 trials per subject, which I guess is not enough. 

1-How many trials should I have per bin (the minimum)?
2-Could I consider the 4 runs as a single run to have more trials?

Any other ideas are welcome.

Thanks again for your work and this forum.

Best wishes,

ane

[n.n.oosterhof]

Greetings,

On Friday, January 12, 2024 at 9:17:19 AM UTC+1 A. G. wrote:

first of all, thanks a lot for such a nice tool!

Thank you.

 

I am new to the cosmo universe,and I was wondering the minimum number of trials I need to obtain reliable results. I run a memory paradigm with 4 conditions and 4 runs (so 16 bins). However, in some bins I only have 3 trials per subject, which I guess is not enough. 

1-How many trials should I have per bin (the minimum)?

That's a difficult question. Generally the answer will include "it depends". On the timing of the design (for example block versus event-related), the estimated difference in BOLD response between conditions, how many voxels are expected to contain information that discriminates between conditions, and probably a few others. That said, 3 trials per subject sounds quite low, and I would advice to think hard if you can change the design to increase the number of trials. 

 

2-Could I consider the 4 runs as a single run to have more trials?

A very common technique is take-one-run-out-crossvalidation. I can imagine that could be used here as well. In any case I think you would want to use some cross-validation (See also https://cosmomvpa.org/ex_classify_double_dipping.html). When using only a few trials per condition, a simple split-half correlation analysis may be a good to start with (https://cosmomvpa.org/ex_splithalf_correlations.html).