Support liberal mask in IBMA estimators

[JulioAPeraza]

Closes None. Related to neurostuff/PyMARE#44

Changes proposed in this pull request:

• Add _apply_liberal_mask to separate the data in bags of voxels in the same studies. Some parts of the proposed function require nested for loops, which was inefficient, so I have adapted some of the procedures to use numba with nopython=True.
• Currently, we loop over the bags and create a Pymare dataset. In the future, we can parallelize that loop to improve performance.

[JulioAPeraza]

Some preliminary results using Stouffers:

[tsalo]

That's so awesome! Do you have a DOF map among the outputs?

[JulioAPeraza]

We don't have them now, but that will be a great addition.

[codecov]

Codecov Report

Attention: 30 lines in your changes are missing coverage. Please review.

Comparison is base (a1c0414) 89.13% compared to head (6dd2e12) 89.10%.

Files	Patch %	Lines
nimare/meta/utils.py	9.09%	30 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #848      +/-   ##
==========================================
- Coverage   89.13%   89.10%   -0.03%     
==========================================
  Files          49       49              
  Lines        6155     6417     +262     
==========================================
+ Hits         5486     5718     +232     
- Misses        669      699      +30     

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

[JulioAPeraza]

Thanks @tsalo! Would that look like the figure below?

[tsalo]

I can't access the linked figure, unfortunately.

[JulioAPeraza]

Sorry, I edited the message yesterday, so the figure probably got deleted.
Could you try refreshing the browser (or accessing the PR without the link in the email) and see if the figure shows in the comment above?

[tsalo]

I can see it now. It looks awesome!

[jdkent]

just a couple changes I requested, but otherwise LGTM, will think of possible refactor of some of the class specific bag handling to be placed in the base class.

[jdkent]

set the number 2 to a variable here, like MIN_STUDY_THRESH, defined at the top of the function.

[jdkent]

I couldn't select all the lines I needed to, but this can be simplified since aggressive_masking is being turned on/off for every IBMA, you can specify another parametrize decorator (see at the bottom, and the ids just name the tests more clearly).

@pytest.mark.parametrize(
    "meta,meta_kwargs,corrector,corrector_kwargs,maps",
    [
        pytest.param(
            ibma.Fishers,
            {},
            FDRCorrector,
            {"method": "indep", "alpha": 0.001},
            ("z", "p", "dof"),
            id="Fishers",
        ),
        pytest.param(
            ibma.Stouffers,
            {"use_sample_size": False},
            None,
            {},
            ("z", "p", "dof"),
            id="Stouffers",
        ),
        pytest.param(
            ibma.Stouffers,
            {"use_sample_size": True},
            None,
            {},
            ("z", "p", "dof"),
            id="Stouffers_weighted",
        ),
        pytest.param(
            ibma.WeightedLeastSquares,
            {"tau2": 0},
            None,
            {},
            ("z", "p", "est", "se", "dof"),
            id="WeightedLeastSquares",
        ),
        pytest.param(
            ibma.DerSimonianLaird,
            {},
            None,
            {},
            ("z", "p", "est", "se", "tau2", "dof"),
            id="DerSimonianLaird",
        ),
        pytest.param(
            ibma.Hedges,
            {},
            None,
            {},
            ("z", "p", "est", "se", "tau2", "dof"),
            id="Hedges",
        ),
        pytest.param(
            ibma.SampleSizeBasedLikelihood,
            {"method": "ml"},
            None,
            {},
            ("z", "p", "est", "se", "tau2", "sigma2", "dof"),
            id="SampleSizeBasedLikelihood_ml",
        ),
        pytest.param(
            ibma.SampleSizeBasedLikelihood,
            {"method": "reml"},
            None,
            {},
            ("z", "p", "est", "se", "tau2", "sigma2", "dof"),
            id="SampleSizeBasedLikelihood_reml",
        ),
        pytest.param(
            ibma.VarianceBasedLikelihood,
            {"method": "ml"},
            None,
            {},
            ("z", "p", "est", "se", "tau2", "dof"),
            id="VarianceBasedLikelihood_ml",
        ),
        pytest.param(
            ibma.VarianceBasedLikelihood,
            {"method": "reml"},
            None,
            {},
            ("z", "p", "est", "se", "tau2", "dof"),
            id="VarianceBasedLikelihood_reml",
        ),
        pytest.param(
            ibma.PermutedOLS,
            {"two_sided": True},
            FWECorrector,
            {"method": "montecarlo", "n_iters": 100, "n_cores": 1},
            ("t", "z", "dof"),
            id="PermutedOLS",
        ),
    ],
)
@pytest.mark.parametrize("aggressive_mask", [True, False], ids=["aggressive", "liberal"])

[jdkent]

There is some potential repetitive code, but that can be refactored later, LGTM!