Skewed Student-T distribution

[fonnesbeck]

Description

Added skewed Student T distribution (Jones and Faddy implementation).

Checklist

• Checked that the pre-commit linting/style checks pass
• Included tests that prove the fix is effective or that the new feature works
• Added necessary documentation (docstrings and/or example notebooks)
• If you are a pro: each commit corresponds to a relevant logical change

Type of change

• New feature / enhancement
• Bug fix
• Documentation
• Maintenance
• Other (please specify):

---

📚 Documentation preview 📚: https://pymc--7252.org.readthedocs.build/en/7252/

[ricardoV94]

Add in pymc-experimental first?

[jessegrabowski]

If something like this goes in experimental, there should be a clear roadmap to go from there back to the main repo. This is a straight-forward distribution that (I guess?) can't be easily implemented as a generative graph using only existing distributions via CustomDist. If it's in pmx, what's the criteria to move it over? It's not something like statespace or marginalmodel that proposes a huge new functionality, or something like r2d2m2 that implements a distribution that doesn't fit neatly into a typical pymc model as we present them in books/tutorials.

[ricardoV94]

I don't think this one has a straightforward generative graph

[ricardoV94]

I assumed something "according to x and y" was a bit experimental, hence my suggestion

[fonnesbeck]

The "Following Jones and Faddy 2003" was to distinguish it from the range of available skewed Student T implementations (there are 3 or 4). I chose this one because it is available in SciPy and relatively straightforward to implement.

I think it will be a useful distribution because it allows one to specify a likelihood that is both skewed and overdispersed and converges to normal when a and b are both large and equal.

[fonnesbeck]

Jax failure does not seem to be related to this PR

[ricardoV94]

Jax failure does not seem to be related to this PR

It's not. Has been failing since the last scipy

[ricardoV94]

But this one is: https://github.com/pymc-devs/pymc/actions/runs/8682226852/job/23806440256?pr=7252#step:7:616

You should try running those logp/logcdf/icdf with n_samples=-1 or whatever it is, to test all combinations locally instead of only a random subset

[ricardoV94]

You're missing the tests for the RV itself. That's also where we cover alternative parametrizations. Something like these

pymc/tests/distributions/test_continuous.py

Lines 2179 to 2200 in 34c2d31

	class TestBeta(BaseTestDistributionRandom):
	pymc_dist = pm.Beta
	pymc_dist_params = {"alpha": 2.0, "beta": 5.0}
	expected_rv_op_params = {"alpha": 2.0, "beta": 5.0}
	reference_dist_params = {"a": 2.0, "b": 5.0}
	size = 15
	reference_dist = lambda self: ft.partial(clipped_beta_rvs, random_state=self.get_random_state()) # noqa E731
	checks_to_run = [
	"check_pymc_params_match_rv_op",
	"check_pymc_draws_match_reference",
	"check_rv_size",
	]
	class TestBetaMuSigma(BaseTestDistributionRandom):
	pymc_dist = pm.Beta
	pymc_dist_params = {"mu": 0.5, "sigma": 0.25}
	expected_alpha, expected_beta = pm.Beta.get_alpha_beta(
	mu=pymc_dist_params["mu"], sigma=pymc_dist_params["sigma"]
	)
	expected_rv_op_params = {"alpha": expected_alpha, "beta": expected_beta}
	checks_to_run = ["check_pymc_params_match_rv_op"]

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 92.35%. Comparing base (60a6314) to head (bcb1b5d).
Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #7252      +/-   ##
==========================================
+ Coverage   91.67%   92.35%   +0.68%     
==========================================
  Files         102      102              
  Lines       17017    17069      +52     
==========================================
+ Hits        15600    15764     +164     
+ Misses       1417     1305     -112     

Files	Coverage Δ
pymc/distributions/__init__.py	100.00% <ø> (ø)
pymc/distributions/continuous.py	97.93% <100.00%> (+0.09%)	⬆️

... and 4 files with indirect coverage changes

[ricardoV94]

Can't sigma and mu be passed as arguments to the rvs?

The way you wrote it has a similar bug that was addressed by this recent bugfix, when there are batched sigma and mu and size is None: #7288

[ricardoV94]

You marked as resolved but I still see the old code?

[fonnesbeck]

Sorry, I had forgotten to push.

[michaelosthege]

Looks like test_skewstudentt_logp is failing on main, or is at least flaky?

[ricardoV94]

These tests run a random subset of 100 combinations. We should set n_samples=-1 locally and see if every combination passes (both float64 and float32), I guess it does not.

CC @fonnesbeck

[michaelosthege]

Running float64 this gives:

E           AssertionError: 
E           Arrays are not almost equal to 6 decimals
E           {'a': array(0.01), 'b': array(100.), 'mu': array(-2.1), 'sigma': array(0.01), 'value': array(-0.01)}
E           x and y -inf location mismatch:
E            x: array(-751.339449)
E            y: array(-inf)

[ricardoV94]

Which underflows? PyMC or Scipy?

[michaelosthege]

Which underflows? PyMC or Scipy?

SciPy produces the -inf.

What's the fix? I have it open and can push a branch real quick (if it's simple)

[ricardoV94]

Which underflows? PyMC or Scipy?

SciPy produces the -inf.

What's the fix? I have it open and can push a branch real quick (if it's simple)

Try slightly less extreme parameter domains