I'd like to conduct a ROI-based analysis in native space and look at the interaction between two within-subject factors (condition (2 levels), accuracy (2 levels)). I am tempted to try out RBA in particular, but I am coming up against a bigger problem of varying trial numbers across my four bins (due to the accuracy factor).
I would imagine LME would be the way to go given the within-subject factors and and varying trial numbers, but given that the analysis has to be in native space I'm gravitating towards methods that can work with mean ROI values, but was wondering whether there was a way to incorporate the varying trial numbers going into the means as a sort of "weight" whether bayesian or frequentist?
Yes, region-based analysis using
RBA can accommodate the uncertainty arising from an unequal number of trials. To clarify, could you confirm if you have effect estimates available at the trial or task level?
Readily available I only have condition level betas e.g.: condition1_accuracy0; condition1_accuracy1, condition2_accuracy0, condition2_accuracy1..
I could attempt to estimate trial level betas although I was initially hesitant as I'm not sure how to estimate the quality of the trial level estimate... With condition level betas I think it's more straightforward..
Can RBA support both approaches?
For each contrast (e.g., condition1 - condition2), obtain its condition-level estimate from each individual, and feed it to
RBA as input.
If you are concerned about the varying number of trials across conditions or individuals, extract the region-level EPI time series and estimate each contrast using
3dDeconvolve at the individual level. Then use both effect estimate and their t-statistic value (or standard error) as input for
So one would separately run an RBA for the main effect of the condition, the accuracy and the interaction, yes?
relatedly: Is there a way to incorporate SE in a non-bayesian analysis to emulate what RBA is doing using the -SE argument?
one would separately run an RBA for the main effect of the condition, the accuracy and the interaction, yes?
Yes, that's correct. This separation is primarily for computational efficiency rather than conceptual or implementational considerations.
Is there a way to incorporate SE in a non-bayesian analysis to emulate what RBA is doing using the -SE argument?
You can utilize
t-statistic as input. This approach allows you to include SE in a manner similar to