Add support for publication size weighting. General refactoring of IBMA estimators

[JulioAPeraza]

Closes None.

Changes proposed in this pull request:

• Add for publication size weighting with a new parameter use_group_size.
• In addition, we perform a general refactoring of IBMA estimators to remove duplicate code and reduce the memory in the estimator.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 88.19%. Comparing base (d95c18b) to head (8fc8030).
Report is 14 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #887      +/-   ##
==========================================
+ Coverage   88.17%   88.19%   +0.02%     
==========================================
  Files          48       48              
  Lines        6374     6335      -39     
==========================================
- Hits         5620     5587      -33     
+ Misses        754      748       -6     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jdkent]

LGTM, just a couple details

I'm trying to think if there would be a case when a contrast wouldn't be included in the actual meta-analysis (e.g., didn't have the correct images for that particular algorithm, or was otherwise filtered from the analysis), but the re-weighing would correct the values as though that contrast was included.

[jdkent]

I'm having second thoughts about use_group_size, group may be more ambiguous than what we are using the variable for. group could mean:

1. patient/control group
2. a group of papers about shizophrenia versus depression

whereas we are using group to mean specifically the number of analyses (contrasts) within a study.

use_group_size could be renamed to normalize_contrast_weights

and group_size to num_contrasts

and group to contrast_names

I think that is more aligned with what the variables are doing.

unless the labels serve more functions than identifying contrasts.

[JulioAPeraza]

I 100% agree! I will make the changes in a bit.

[JulioAPeraza]

I'm trying to think if there would be a case when a contrast wouldn't be included in the actual meta-analysis (e.g., didn't have the correct images for that particular algorithm, or was otherwise filtered from the analysis), but the re-weighing would correct the values as though that contrast was included.

In the new _preprocess_input in Stouffers, the number of contrasts is estimated from self.inputs_["id"], so any excluded images will also be excluded from the count.

[jdkent]

LGTM, thanks @JulioAPeraza!