TEST: dataset-level model spec retrieval

[tyarkoni]

Very minor test that verifies you can retrieve a model specification object for a BIDS-StatsModels analysis at levels past run (in this case, dataset). (I think black went a little crazy on the formatting, which was unintended, but the only substantive changes are the addition of test_get_dataset_level_model_spec.)

[effigies]

Rebased to run tests.

[effigies]

@tyarkoni I'm going to revert all style changes except test_get_dataset_level_model_spec, since there are errors in unrelated tests.

[tyarkoni]

Sounds good, sorry about that.

[effigies]

Actually, I think the failures are related to the change in the model. Never mind.

[tyarkoni]

Ah, sorry—I forgot to run the other tests before pushing.

These failures actually suggest a deeper problem that will probably require more than just amending the tests. It looks like pybids doesn't currently broadcast higher-level metadata properly when there's more than one row per unit passed from the previous subject. E.g., if you go straight from run to dataset level, and you're passing 10 runs from 2 subjects, then at the dataset level you have 10 rows, but participants.tsv only contains 2 rows, so you need to broadcast the latter to have the same shape as the former. Will fix this when I can.

[tyarkoni]

@effigies tests are breaking because of #682; let's ignore that here and I'll deal with it separately (I think the issue is that the JSON changes @oesteban introduced to cause #683 and trigger the new test are breaking other things).

If you want to review this, the key changes are in this section. My earlier assessment was wrong; the issue here isn't broadcasting-related, it's that bad things happen when metadata is missing for some rows but not others (e.g., the rows propagated to the dataset level from runs --> sessions --> subjects carries information about the scanning sequence, but this is missing from rows ingested from participants.tsv). Things were already working fine in "long" format but were breaking in "wide"; I've now fixed it so it works both ways.

[oesteban]

Seems to be rooting from the new bids/tests/data/7t_trt/sub-01/ses-1/sub-01_ses-1_scans.json file. If doing that kind of thing is explicitly ruled out in the specs, that file can just be deleted.

I wonder if it would suffice to first merge #683 and use the hashablefy function on line 689 of the layout code where things break. Isn't #683 updated so that this error can be replicated there?

[tyarkoni]

@effigies we should probably merge this before mucking around further with the design matrix construction stuff... the test failures look like a CI glitch, and are otherwise passing.