I’m trying to a single trial pattern analysis trying to derive beta estimates from events that are very close to each other in time. One event follows the next immediately so I got an event at 30s then at 32s and at 34s. I’m using the least square all (LSA) approach and have implemented this in nistats. However, I’m getting strange results, I’m wondering whether this may be due to the fact that the events are so close to each other in time?
Is there a work around for that? Would it help to run separate model for each of the events instead of estimating them all in one model?
Thanks for your help,
LSS is designed for rapid event-related tasks, although it still requires jittering to make sense. Based on your description of your task, it was unclear to me whether there is jittering or not.
@bthirion has made solid arguments against LSS in favor of LSA in Convert beta estimates to t-statistics for whole brain searchlight MVPA - is it necessary for my task?, although I’m not sure if those are specific to (1) tasks with non-overlapping HRFs or (2) the old version of LSS (in which non-target conditions are combined).
I still lean toward using LSS, in which the one model is run for each event, with that event “popped out” into its own condition but everything else remaining the same. However, I have only used beta series models for functional connectivity in the past and I am not experienced with MVPA, so take my recommendation with a grain of salt.
If you have the time to try it out, you might want to consider the method from Loula, Varoquaux, & Thirion (2018), which outperforms LSS for short ISIs in their analyses. Here is the code: https://github.com/Joaoloula/time_decoding
I haven’t used that tool, but it seems promising and hopefully someone who knows more about it can help you out, if you choose to use it.
If you can share an example (including design matrix and weird result) we might be able to provide some feedback.
It may indeed be the case that the events are too close in time, so that the model hrf are strongly overlapping, compromising parameter identification.