Application to voxel-wise data

ModelArray is compatible with not only fixel-wise data, but also voxel-wise data. The vignette("walkthrough") page has shown example walkthrough with example fixel-wise data, with hints for voxel-wise data provided throughout the walkthrough. In general, the workflow of applying ModelArray to voxel-wise data is consistent to that of fixel-wise data, but please note several important differences when execution described as below.

Conversion of voxel-wise data format

When converting voxel-wise data between NIfTI and HDF5 file format that ModelArray requires, you will still use ConFixel software, but a different converter tailored for voxel-wise data, ConVoxel. We have provided a walkthrough of how to use it here.

Applying ModelArray to voxel-wise data

You can apply the same ModelArray functions, arguments, etc used in fixel-wise data to voxel-wise data; just be aware that an “element” is a “voxel” now.

In addition, there are additional arguments specifically useful for voxel-wise data, regarding the subject-specific masks (i.e., individual masks). Different from fixel-wise data where subjects usually share the same mask, voxel-wise data may have subject-specific masks, where different subjects may have different brain coverage even in template space. Therefore, for voxels (e.g., at the edge of the brain) which do not have data from sufficient number of subjects, their statistic results might not be reliable. It’s better to exclude those voxels from the analysis.

In ConVoxel, for voxels not within a subject-specific mask (but within a larger, group-level mask), their values will be set to NaN. Thus ModelArray can know the number of subjects that have valid values in each voxel within the group mask.

To set up minimal number of subjects required for a voxel, you can use these two arguments in ModelArray.lm() and ModelArray.gam():

• num.subj.lthr.abs: lower threshold, absolute number of subjects; e.g., 10
• num.subj.lthr.rel: lower threshold, proportion of total number of subjects, a value between 0-1; e.g., 0.2 (i.e., 20% of all subjects)

Suppose a voxel has valid values from N subjects, where N:

• N > num.subj.lthr.abs, and
• N > num.subj.lthr.rel * Ntotal

where Ntotal is the total number of subjects, which is quantified by number of rows in argument phenotypes. If N satisfied above criteria, statistical analysis of this voxel will be performed as normal; otherwise, this voxel will be skipped and statistical outputs will be set as NaN.

For more details about these two arguments, please check the reference pages of ModelArray.lm() and ModelArray.gam().

Finally, if you choose to save number of observations used (i.e., if specifying nobs in var.model in ModelArray.lm() and ModelArray.gam()), after conversion using ConVoxel, you’ll see an image called *_model.nobs*. If the subject-specific masks you provided are different across subjects, in this image you’ll probably see the heterogeneity of number of observations used in the brain, with fewer observations used at the edge of the brain.