Hi Rick,

Another issue I was wondering about was how to evaluate how well my model fits the data after I have run 3dDeconvolve. I'm more familiar with the time series plots produced by FSL's FEAT analysis, where it is as simple as seeing whether the model and data line up relatively well. I'm sure there must be a way to produce a similar graph in AFNI. I see there are the -fitts and -errts output files, which I would think could be used to evaluate fit and error but not the model itself.

To answer your question, I had deleted the first 17 TRs (which included the first breathhold) from the input file. The task was mapped out in Excel so that the remaining breathhold onset times were accurately adjusted to the following: 7, 54, 101, 148, 195. I used these as the stimulus onset times with a block (15,1) function. However, when I did this, the resulting stats output viewed in AFNI showed lower activation associated with the breathhold. This suggested to me that there must be something wrong with the regression.

For the breathhold task, some individual's hemodynamic response peak relatively early after the end of the breathhold; others take longer. But in either case the rise and fall of the hemodynamic response does not fit cleanly within the 15s of the breathhold. So, it's necessary to shift the model to the right to fit the model well to the data. Some analyses take the approach of using the same delay (often 9s) for all subjects to accommodate this delay in the hemodynamic response. In our analyses we used an individualized delay value for each subject that is the average time required for the BOLD response to peak after the end of the breathhold for that subject. I tried adding this individualized delay value to each of the breathhold onset times and it resolved the direction issue. The breathhold is now associated with increased activation. However, I'm concerned that I have simply fit the hemodynamic response function to start at the peak of the BOLD response in the data.

Ultimately, my main questions boil down to: 1) in trying to adjust my model, how do I account for the specific delay times applied by BLOCK or the other techniques (do these functions try to fit the hemodynamic response within the 15s duration I have specified, or do they also shift to the right by some number of seconds); 2) how do I adjust my model to fit the data well, and 3) how do I check that the model fits the data well? I hope this makes sense.

Edit: After posting this message, I found this (https://afni.nimh.nih.gov/sscc/gangc/Fit.html), and used the second technique to evaluate the fit of the regression. It appears that the red curve is somewhat left of the original signal.

Edit 2: If my understanding is correct, this is the fit of the model rather than the model itself. Is there a way to plot just the model against the data? Also, is there a way to do this at a whole brain level rather than voxel-by-voxel, similar to the timeseries plots produced by FSL's FEAT analysis?



Edited 5 time(s). Last edit at 05/03/2016 10:33AM by Alain.