Raise warning if coordinates dataset contains both positive and negative z_stats

[jdkent]

Summary

Seems weird to perform a meta-analysis with peaks being reported at both the positive and negative thresholds

Additional details

• NiMARE version: 0.0.7 dev

What were you trying to do?

Thinking about how users may use the output of #446

What did you expect to happen?

Raise warning/error when both positive/negative values are detected in z_stat column?

What actually happened?

The dataset will run on any estimator

Reproducing the bug

[tyarkoni]

It's definitely less common than using unidirectional coordinations, but there is a use case for it—e.g., if you have a bunch of contrasts where there's no clear directional hypothesis, but you think a consistent set of regions should be implicated in some way. Since density-based analyses don't care about sign, you will still get sane results, provided you adjust your interpretation appropriately. Neurosynth, for instance, works exclusively like this (though that's more by necessity than design).

So, with that in mind, I don't think we want to raise an error. But a warning does seem appropriate.

[jdkent]

Thanks, great point, the interpretation is the crucial step here, not the analysis itself.

[tsalo]

I would default to splitting the contrasts within the Dataset during the img-to-coordinate conversion, with a warning. If you wanted to be really fancy, you could also invert the images and save them out as well. So study X's contrast of "A>B" would have positive z-statistic coordinates, and then the conversion would create a new contrast, "B>A" with negative z-statistic coordinates and possibly the inverted images.

EDIT: Of course, things might change if we ever get an algorithm that works with effect sizes and bidirectional hypotheses, like SDM, working!

Ultimately, though, it seems like a NIMADS/Dataset-searching problem to me. We don't really take into account sign/contrast definition, even with images, in our current approach...

[tsalo]

Just to make sure I understand the current state of this- we do want to raise a warning, right?

[jdkent]

I think a warning is appropriate.