meta-analysis problem

[adelavega]

summary: afni meta-anaysis maps w/ all tasks look very weak, especially for RMS which should be solid and looks good at single dataset level (t-stats)

two possibilities: 1) nndb is on a different scale from other datasets, due to preprocessing differences (primarily) but also predictor scale differences (q: should we re-scale?)

• look at time series for both
• possibly rescale RMS for an NNDB task and take a look 2) nndb tasks on average have fewer subjects, and if that's a problem (i.e. higher variance), they may overly domiante meta-analysis 3) AFNI specific issue may lead to unstable PEs, with some very high value
• find an example of this
• exclude NNDB from meta-analysis to test if that's the primary cause of strange results

[jdkent]

Notes:

• nistats has not been run on nndb so the results should be identical

• using nistats by itself does not have strong meta-analytic maps since estimator has not always been tracked (but I could assume anything without an estimator was run using nistats).

• afni:

  • nndb:
  • yes #############################:
  • no #############################:

• nistats

  • nndb:
  • yes #############################:
  • no #############################:

[jdkent]

nistats including "None" estimators (assuming they are nistats)

[jdkent]

using the first available collection for every study (including nndb )

[jdkent]

first available collections without nndb

[jdkent]

the first available collection results with and without nndb are perfectly correlated, but differ in magnitude, where not including nndb increases magnitude.

[adelavega]

Looks most likely that the issue is with fitlins. https://github.com/poldracklab/fitlins/pull/318

Running a test model to see if the patch fixes the issue. Analysis ID: MzQpZ

If it seems to help, then run on all non-NNDB datasets, and compare to the map above. If so, will delete old maps, with afni since they are incorrect.

[jdkent]

problem is neurostuff/NiMARE/issues/579