Dear AFNI experts,
In trying to prep data for an RSA analysis, your 3dLSS program is wonderfully efficient. However, in addition to beta-coefficient images, I’d also like to have t-statistic maps to get a metric that also includes info on variability. As far as I can tell from the help, 3dLSS does not have an option for outputting t-maps. Is there a way to have it do this that I’m just not seeing? Or will I just need to run 3dDeconvolve for each stimulus for each subject in the experiment like in the bad old days before 3dLSS was available?
In trying to prep data for an RSA analysis, your 3dLSS program is wonderfully efficient.
Efficient in what sense? Compared to 3dDeconvolve/3dREMLfit?
3dLSS does not have an option for outputting t-maps.
From the modeling perspective, 3dLSS is inferior to 3dDeconvolve/3dREMLfit. Essentially 3dLSS is a band-aid approach to fixing the potential numerical problem of high collinearity when the individual trials are too close to each other. If your experimental design is well-optimized, it would be much preferable to directly use 3dDeconvolve/3dREMLfit instead of 3dLSS. It is also because of the band-aid approach with 3dLSS that makes it difficult to provide statistical evidence (e.g., t-statistics).
Thanks for the helpful reply! To be frank, I hadn’t thought about 3dREMLfit. My near-term goal was to get a volume of beta-weight estimates and corresponding t-scores for each stimulus event in an event-related design (for later use with RSA). Before 3dLSS existed the only way I knew to do that was to run 3dDeconvolve once for each stimulus, modeling all other stimuli separately. Just using 3dDeconvovle to generate the design matrix and then feeding that to 3dLSS to make the maps for each stimulus is orders of magnitude faster, so that’s what I was referring to.
But now that you mention it, if I read the program help correct, it looks like 3dREMLfit will also generate a map per stimulus without having to run it separately for each stimulus, AND it will output nice t-maps as well. If that’s right, then it definitely seems worth a shot!
Before 3dLSS existed the only way I knew to do that was to run 3dDeconvolve once for each stimulus, modeling all other stimuli separately.
Unless your trials are really too close to each other, why not use option -stim_times_IM in 3dDeconvolve and model the trials simultaneously (instead of separately)?
Ha, because I didn’t know the “-stim_times_IM” option in 3dDeconvolve worked that way until just now. But before I learned that from your most recent reply, I used 3dREMLfit as a drop-in replacement for 3dLSS like you suggested, and it seemed to work beautifully. So thank you Gang!