Global rng

[amichuda]

resolves #104

• Adds ability to add a numpy.random.Generator as argument to seed
• Changes function description and adds type hints in line with above
• Adds test to check that adding integer seed and generator leads to same answers

[amichuda]

Rather than adding a new argument, I just went ahead and allowed seed to take a random generator as an input. I think this is better than adding a new argument as it cuts down on the need for extra logic or exception handling if both seed and a hypothetical rng argument would be both specified. Let me know if that works

[s3alfisc]

Looks good and makes a lot of sense!

[amichuda]

@s3alfisc Great, thank you! Is there a way to re-run the CI checks now that the github action is updated?

EDIT: nvm, figured it out!

[amichuda]

So it looks like the tests for seeds are failing, but they worked yesterday, I'm going to make the tests a bit more robust and try multiple seeds.

[amichuda]

@s3alfisc

So as it stands, this part of the test_results_from_same_seed fails:

# random seed outside of function 2x -> same results
np.random.seed(123)
a2 = wildboottest(model, param = "X1", cluster = x, B= 999)
np.random.seed(123)
b2 = wildboottest(model, param = "X1", cluster = x, B= 999)
pd.testing.assert_frame_equal(a2,b2)

Right now, if you don't add a seed explicitly, wildboottest creates a default_rng with seed=None so it's not reproducible and is also not changed if the seed is changed globally.

Is setting the random seed globally like this a big use-case for py-econometrics? If so, I can work on setting it inside wildboottest, but note that numpy does not recommend this behavior:

https://numpy.org/doc/stable/reference/random/generated/numpy.random.seed.html