search

incf-nidash / nidm-specs

Neuroimaging Data Model (NIDM): describing neuroimaging data and provenance
nidm.nidash.org
Other
33 stars 30 forks source link

Potential new fields in NIDM-Results? #170

Closed cmaumet closed 10 years ago

cmaumet commented 10 years ago

Potential info to be added in NIDM-Results, discussed with @gllmflndn:

tiborauer commented 10 years ago

Are they results generated by the software (e.g. FSL, SPM) or by the NIDM-Results "plugin" (i.e. new results)?

nicholst commented 10 years ago

+1 for these changes, as they will quickly become important as we add FSL support. Would be clearer if we came up with some clear description of how these are 'filtering' aspects, i.e., minor (but important) options that express how peaks are extracted from an image. Something like

Then about estimation methods... unfortunately STATO only offers us linear regression for analysis of continuous dependent variable (STATO_0000108).

I would say the following estimation methods (for a general linear model) are relevant and need to be represented:

There are no end of advanced estimation methods (and estimation methods that are inextricably linked to their tests, e.g. Mann-Whitney) that we will not add to the model unless there are specific instances where they are used in neuroimaging.

With either WLS or GLS, a crucial aspect is how the parameters associated with the heteroscedasticity and/or dependence are parameterised and estimated, and I assume this is what @gllmflndn was alluding to by "noise assumptions". For the purposes of neuroimaging, what matters is whether these parameters are assumed common over the whole brain (global) or estimated locally at each voxel, and if the latter, if they have been regularised in some way. The details in the parameterisation, beyond that, get into a worm-hole of detail, but perhaps a "NoiseAssumptions" could take on values like

Again... there are tons of details, but we should discuss what resolution is important to record.

  1. Wager TD, Keller MC, Lacey SC, Jonides J. Increased sensitivity in neuroimaging analyses using robust regression. Neuroimage. 2005;26(1):99–113.
cmaumet commented 10 years ago

Noise models and estimation methods are now implemented in #176.

cmaumet commented 10 years ago

@tiborauer: sorry for the late reply. The results that we are modelling in NIDM-Results are the one created by the analysis software (SPM, FSL). NIDM-Results plugins (natively in SPM12b and as Python scripts for FSL) are used to export existing results in the NIDM-Results "format". Does that makes more sense?

tiborauer commented 10 years ago

@cmaumet: yes. Thank you for returning to me! For the second two, it is clear (Description of the methods). However, the first two seems to me estimates which are not stored anywhere by the original program, but estimated by the NIDM-Results plugin (i.e. new results). It just goes back to a previous (probably oudated) discussion about whether we should generate new results.

cmaumet commented 10 years ago

@tiborauer: I agree, I think this relates to the discussion with had in #114. NIDM-Results is special in the sense that we are recording information in a (as much as possible) software-agnostic fashion.

However, I would not necessarily make a distinction between those four pieces of info. All of them are recorded in some ways by the analysis software. As you mentioned "noise models" and "estimation methods" are descriptive of the method used. "maximum number of peaks per cluster" and "minimal distance between peaks" are pre-defined parameters that can also be retrieved (e.g. in SPM at the bottom of the results page "table shows xx local maxima more than xx mm apart").

tiborauer commented 10 years ago

@cmaumet: You are right! Thank you for the clarification. It was not clear to me whether these parameters are pre-set (to determine local maxima) or estimated (to describe local maxima).

cmaumet commented 10 years ago

Creation of those fields are now under discussion in separate pull requests: