Hi Jacob, 

whenever there is is systematic difference, there is also potential for a bias. In your specific situation no one knows how large this bias could be. I would think it will not really be measurable, as the base is only used for initialisation and within such a small time frame, the methods should be flexible enough to fine-tune to the final time points, independent of where the initialisation starts. So I would not be worried too much. I also don’t know the numbers of your situation. E.g. if the majority of subjects has 4 time points and only some are missing one in the middle, I would be worried even less. 

The only concern would be when lots of subjects miss one of the middle time points and when there is a strong bias across your variable of interest (e.g. disease group). For example, most healthy are missing time point 3 and most disease are missing time point 2. That setting could be problematic (and as I said above, even that setting probably is not problematic). 

The weighting you describe does not really work. To increase robustness we are using a voxel-wise intensity median to create the base image. 

If you are really concerned (e.g. because of a strong bias across groups), you could create the base from the first and last time point only, then add the two or one additional middle time points to that base later. That way all are processed in a similar way. I once created an -add-tp option for adding time points later without re-processing base and all longs. 

Best, Martin 


On 14. Dec 2021, at 04:24, Jacob Levenstein <jacob.levenstein@ndcn.ox.ac.uk> wrote:


Dear Freesurfer friends,  

 

It seems that prior messages* on longitudinal recon with missing timepoints have suggested that it is okay for a subset of participants to include different numbers of timepoints into the longitudinal pipeline, so long as there is not a systematic bias (e.g., one group consistently contains more scans than another group, etc.).  

 

Assuming that is still true, I am working with an adolescent dataset from 11 year-olds scanned every 3 months for a year (4 scans total). For participants missing a scan at either timepoint 2 or timepoint 3, I was wondering if creating the BASE/subject-specific template from only three sessions might introduce a temporal bias toward either timepoint 1 (i.e., formed from timepoints 1, 2 & 4), or bias toward timepoint 4 (i.e., formed from timepoint 1, 3 & 4). With consideration to the developmental changes we anticipant measuring over the year, would you be concerned with the potential early vs late skew in the BASE for these missing data cases? 

 

If so, for the missing timepoint participants, I was wondering if I should then avoid generating their subject-specific templates in a way that imposes an even weighting across the three timepoints (i.e., .33 each), and instead adjust the weighting of each timepoint relative to time acquired over the year (e.g., temporal based average), such that timepoint 1 is weighted at .25 and timepoint 2 and 4 are weighted at .375. If this is not a sensible approach, could you please let me know what you think best? 

 

Thank you for your insight! 

 

Much appreciated,  
Jacob  

 


_______________________________________________
Freesurfer mailing list
Freesurfer@nmr.mgh.harvard.edu
https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer