TestCorrelation::test_corr fails on some platforms

[musicinmybrain]

I’m helping to update the python-pingouin package in Fedora Linux. One test, TestCorrelation::test_corr, is failing on some platforms because what should theoretically be a perfect unity correlation is coming out two or three ulps short:

        # Compare BF10 with JASP
        df = read_dataset('pairwise_corr')
        stats = corr(df['Neuroticism'], df['Extraversion'])
        assert np.isclose(1 / float(stats['BF10'].to_numpy()), 1.478e-13)
        # Perfect correlation, CI and power should be 1, BF should be Inf
        stats = corr(x, x)
>       assert stats.at['pearson', 'r'] == 1
E       AssertionError: assert 0.9999999999999997 == 1

This happens when the tests are executed on an aarch64, ppc64le, or s390x platform. (On ppc64le and s390x, the correlation is 0.9999999999999998.)

What do you think? Does this look like this a bug to be fixed somewhere, or should the assertion be loosened to permit a value slightly less than one due to platform-dependent rounding?

[raphaelvallat]

Hi @musicinmybrain,

Interesting — thanks for opening the issue and helping with this! I think it is definitely a platform-dependent rounding error so we should use np.isclose(stats.at['pearson', 'r'], 1) instead of the strict ==.

Raphael

[musicinmybrain]

Makes sense to me. You could, if you like, add an additional test for stats.at['pearson', 'r'] <= 1', since a high-quality implementation shouldn’t be producing correlations outside [-1, 1] no matter how the rounding happens to work out.

[raphaelvallat]

Done in 7f5f0cc

See PR v0.5.1: #236