Mass univariate analysis and spatially varying design matrix

[cmaumet]

This is a proposal made up with @nicholst, following discussions at #191 to:

• specify whether the analysis is "mass univariate" by including a new attribute nidm:hasVoxelWiseEstimation (or nidm:hasPixelWiseEstimation) in nidm:ModelParametersEstimation that take as value an nidm:ElementWiseEstimation:

• specify whether the design matrix is spatially varying by re-using the attribute dependenceSpatialModel in nidm:DesignMatrix that take as value an nidm:SpatialModel (terms under discussion at Definition of "noise spatial model" terms #194):

[nicholst]

I think some exposition and/or examples are needed to clarify what's going on here.

Here are what the values nidm:hasVoxelWiseEstimation can take on mean (in a more logical order, instead of alphabetical)

• MassUnivariate: This is the traditional, ubiquitous approach; parameter estimation depends solely on data at each individual voxel/pixel/element.
• MassMultivariate: Less common approach, where a multivariate data is available at each voxel/pixel/element (e.g. Tensor Based Morphometry, or a 3x3 tensor from DTI); (multivariate) parameter estimation depends solely on (multivariate) data at each individual voxel/pixel/element.
• LocalMultivariate: Recently popularised "spotlight" approach, where a local, spatial multivariate model is fit at each location, centered about each individual voxel/pixel/element, using a neighbourhood of observations.
• Multivariate: Classical, but infrequently used approach, where the entire image is taken as a single multivariate observation; (multivariate) parameter estimation depends on the entire collection of voxel/pixel/element over the analysis mask.

For dependenceSpatialModel in nidm:DesignMatrix, I'm not sure a SpatiallyRegularisedModel makes sense for the design matrix. That is, the design matrix is a fixed and given object; "regularisation" implies a modification of an estimation procedure, which doesn't really make sense in the context of design matrices.

[cmaumet]

We discussed this on NIDASH call on Nov 3rd (minutes).

[khelm]

Just a couple of comments:

1. The definition should start with what the thing is: MassUnivariate: "A model in which the estimation of the model parameters depends solely on the data at each individual voxel/pixel/element."
2. I don't think that the commentary on the current community usage is subjective and shouldn't be in the definitions.

[nicholst]

thanks @khelm! Indeed, I wasn't thinking about finalised definitions but trying to make the different terms clear to the NIDM group. I can create revised definitions that are more formal.

[khelm]

Well, I didn't want a call to go by without making at least one nit-picky comment about definitions...
(and strike the "don't" in comment #2 above).