Fixing two_sample default noise_model

[dburkhardt]

default noise_model in the two_sample helper function now matches the docstring and other testing methods

[davidsebfischer]

Thanks for the PR @dburkhardt! I put this as None because t-test complains if it gets a noise model that is not gaussian. This way, t-test can always be run easily and people have to choose (I considered this more advanced) if they use a wald test. Happy for feedback about this design choice! I could also change the docstring to make this clear.

[dburkhardt]

Hmm okay, I think I'm seeing the issue here. So my guess is that for cases where you're comparing two samples of scRNA-seq, "nb" is the correct noise model for the Wald test.

If you're comparing two clusters, I think none of these tests will yield useful p-values because clustering introduces differences between partitions by design (https://linkinghub.elsevier.com/retrieve/pii/S2405471219302698).

What do you think here? If "nb" is the correct noise model for two_sample comparisons (i.e. comparisons of independently generated sets of cells), then why not have that set by default?

[davidsebfischer]

My take on the different tests is that they represent different assumptions on the data distribution and the necessity / way of inclusion of confounding variables. I agree with this

What do you think here? If "nb" is the correct noise model for two_sample comparisons (i.e. comparisons of independently generated sets of cells), then why not have that set by default?

But I would translate it to, if one choses a wald test, then "nb" is set as the default noise model. I can do that internally in the two_sample function. Then the default choise for noise_model is None which is ok for t-tests and which is changed to nb if wald test is chosen. Does match what your intuition?