import pytest import numpy as np from numpy.testing import assert_allclose from scipy.conftest import skip_xp_invalid_arg import scipy._external.array_api_extra as xpx from scipy._lib._array_api import (make_xp_test_case, xp_default_dtype, is_jax, is_cupy, eager_warns, xp_result_type, is_array_api_strict) from scipy._lib._array_api_no_0d import xp_assert_close, xp_assert_equal from scipy import stats from scipy.stats._axis_nan_policy import SmallSampleWarning lazy_xp_modules = [stats] @make_xp_test_case(stats.chatterjeexi) class TestChatterjeeXi: @pytest.mark.parametrize('case', [ dict(y_cont=True, statistic=-0.303030303030303, pvalue=0.9351329808526656), dict(y_cont=False, statistic=0.07407407407407396, pvalue=0.3709859367123997)]) @pytest.mark.parametrize('dtype', ['float32', 'float64', None]) def test_against_R_XICOR(self, case, dtype, xp): # Test against R package XICOR, e.g. # library(XICOR) # options(digits=16) # x = c(0.11027287231363914, 0.8154770102474279, 0.7073943466920335, # 0.6651317324378386, 0.6905752850115503, 0.06115250587536558, # 0.5209906494474178, 0.3155763519785274, 0.18405731803625924, # 0.8613557911541495) # y = c(0.8402081904493103, 0.5946972833914318, 0.23481606164114155, # 0.49754786197715384, 0.9146460831206026, 0.5848057749217579, # 0.7620801065573549, 0.31410063302647495, 0.7935620302236199, # 0.5423085761365468) # xicor(x, y, ties=FALSE, pvalue=TRUE) dtype = xp_default_dtype(xp) if dtype is None else getattr(xp, dtype) rng = np.random.default_rng(25982435982346983) x = rng.random(size=10) y = (rng.random(size=10) if case['y_cont'] else rng.integers(0, 5, size=10)) x, y = xp.asarray(x, dtype=dtype), xp.asarray(y, dtype=dtype) res = stats.chatterjeexi(x, y, y_continuous=case['y_cont']) xp_assert_close(res.statistic, xp.asarray(case['statistic'], dtype=dtype)) xp_assert_close(res.pvalue, xp.asarray(case['pvalue'], dtype=dtype)) @pytest.mark.parametrize('y_continuous', (False, True)) def test_permutation_asymptotic(self, y_continuous): # XICOR doesn't seem to perform the permutation test as advertised, so # compare the result of a permutation test against an asymptotic test. rng = np.random.default_rng(2524579827426) n = np.floor(rng.uniform(100, 150)).astype(int) shape = (2, n) x = rng.random(size=shape) y = (rng.random(size=shape) if y_continuous else rng.integers(0, 10, size=shape)) method = stats.PermutationMethod(rng=rng) res = stats.chatterjeexi(x, y, method=method, y_continuous=y_continuous, axis=-1) ref = stats.chatterjeexi(x, y, y_continuous=y_continuous, axis=-1) np.testing.assert_allclose(res.statistic, ref.statistic, rtol=1e-15) np.testing.assert_allclose(res.pvalue, ref.pvalue, rtol=2e-2) def test_input_validation(self, xp): rng = np.random.default_rng(25932435798274926) x, y = rng.random(size=(2, 10)) x, y = xp.asarray(x), xp.asarray(y) message = 'Array shapes are incompatible for broadcasting.|Incompatible shapes' with pytest.raises((ValueError, TypeError), match=message): stats.chatterjeexi(x, y[:-1]) if not (is_jax(xp) or is_cupy(xp)): # jax misses out on some input validation from _axis_nan_policy decorator # This also fails with cupy for reasons that need to be investigated. message = '...axis 10 is out of bounds for array...|out of range' with pytest.raises((ValueError, IndexError), match=message): stats.chatterjeexi(x, y, axis=10) message = '`y_continuous` must be boolean.' with pytest.raises(ValueError, match=message): stats.chatterjeexi(x, y, y_continuous='a herring') message = "`method` must be 'asymptotic' or" with pytest.raises(ValueError, match=message): stats.chatterjeexi(x, y, method='ekki ekii') def test_special_cases(self, xp): message = 'One or more sample arguments is too small...' with pytest.warns(SmallSampleWarning, match=message): res = stats.chatterjeexi(xp.asarray([1]), xp.asarray([2])) assert xp.isnan(res.statistic) assert xp.isnan(res.pvalue) @make_xp_test_case(stats.spearmanrho) class TestSpearmanRho: @pytest.mark.parametrize('alternative, statistic, pvalue', [ ('two-sided', -0.2727272727272727, 0.4458383415428), ('greater', -0.2727272727272727, 0.7770808292286), ('less', -0.2727272727272727, 0.2229191707714), ]) @pytest.mark.parametrize('dtype', ['float32', 'float64', None]) def test_against_R_cor_test(self, alternative, statistic, pvalue, dtype, xp): # Test against R cor.test, e.g. # options(digits=16) # x = c(0.11027287231363914, 0.8154770102474279, 0.7073943466920335, # 0.6651317324378386, 0.6905752850115503, 0.06115250587536558, # 0.5209906494474178, 0.3155763519785274, 0.18405731803625924, # 0.8613557911541495) # y = c(0.8402081904493103, 0.5946972833914318, 0.23481606164114155, # 0.49754786197715384, 0.9146460831206026, 0.5848057749217579, # 0.7620801065573549, 0.31410063302647495, 0.7935620302236199, # 0.5423085761365468) # cor.test(x, y, method='spearman', alternative='t', exact=FALSE) dtype = xp_default_dtype(xp) if dtype is None else getattr(xp, dtype) rng = np.random.default_rng(25982435982346983) x = rng.random(size=10) y = rng.random(size=10) x, y = xp.asarray(x, dtype=dtype), xp.asarray(y, dtype=dtype) res = stats.spearmanrho(x, y, alternative=alternative) xp_assert_close(res.statistic, xp.asarray(statistic, dtype=dtype)) xp_assert_close(res.pvalue, xp.asarray(pvalue, dtype=dtype)) @pytest.mark.parametrize('alternative, statistic, pvalue', [ ('two-sided', -0.4857142857142857, 0.3555555555556), ('greater', -0.4857142857142857, 0.8513888888889), ('less', -0.4857142857142857, 0.1777777777778), ]) def test_against_R_cor_test_exact(self, alternative, statistic, pvalue): xp = np # will test for multiple backends when gh-23772 merges # Test against R cor.test exact=TRUE, e.g. # options(digits=16) # x = c(0.11027287231363914, 0.8154770102474279, 0.7073943466920335, # 0.6651317324378386, 0.6905752850115503, 0.06115250587536558) # y = c(0.5209906494474178, 0.3155763519785274, 0.18405731803625924, # 0.8613557911541495, 0.8402081904493103, 0.5946972833914318) # cor.test(x, y, method='spearman', alternative='t', exact=TRUE) rng = np.random.default_rng(25982435982346983) x = rng.random(size=6) y = rng.random(size=6) x, y = xp.asarray(x), xp.asarray(y) method = stats.PermutationMethod() res = stats.spearmanrho(x, y, method=method, alternative=alternative) xp_assert_close(res.statistic, xp.asarray(statistic)) xp_assert_close(res.pvalue, xp.asarray(pvalue)) @pytest.mark.parametrize('alternative', ('two-sided', 'greater', 'less')) @pytest.mark.parametrize('n', [9, 99, 999]) def test_against_scipy_spearmanr(self, alternative, n, xp): rng = np.random.default_rng(5982435982346983) x = rng.integers(n//2, size=n) y = rng.integers(n//2, size=n) dtype = xp_default_dtype(xp) ref = stats.spearmanr(x, y, alternative=alternative) ref_statistic = xp.asarray(ref.statistic, dtype=dtype) ref_pvalue = xp.asarray(ref.pvalue, dtype=dtype) x = xp.asarray(x, dtype=dtype) y = xp.asarray(y, dtype=dtype) res = stats.spearmanrho(x, y, alternative=alternative) xp_assert_close(res.statistic, ref_statistic) xp_assert_close(res.pvalue, ref_pvalue) @pytest.mark.parametrize('axis', [-1, 0, 1]) def test_other_backends_nd(self, axis, xp): # NumPy n-d behavior is tested by `test_axis_nan_policy`; # check other backends against it. rng = np.random.default_rng(8243598346983259) shape = (8, 9, 10) x = rng.standard_normal(size=shape) y = rng.standard_normal(size=shape) res = stats.spearmanrho(xp.asarray(x), xp.asarray(y), axis=axis) ref = stats.spearmanrho(x, y, axis=axis) xp_assert_close(res.statistic, xp.asarray(ref.statistic), atol=1e-16) xp_assert_close(res.pvalue, xp.asarray(ref.pvalue), atol=1e-16) def test_input_validation(self, xp): rng = np.random.default_rng(25932435798274926) x, y = rng.random(size=(2, 10)) x, y = xp.asarray(x), xp.asarray(y) msg = 'incompatible for broadcasting|Incompatible shapes|must be broadcastable' with pytest.raises((ValueError, TypeError), match=msg): stats.spearmanrho(x, y[:-1]) if not is_jax(xp): # jax misses out on some input validation from _axis_nan_policy decorator message = '...axis 10 is out of bounds for array...|out of range' with pytest.raises((ValueError, IndexError), match=message): stats.spearmanrho(x, y, axis=10) message = "`alternative` must be 'less', 'greater', or 'two-sided'." with pytest.raises(ValueError, match=message): stats.spearmanrho(x, y, alternative='alternative') message = ("`method` must be...") with pytest.raises(ValueError, match=message): stats.spearmanrho(x, y, method='method') def test_special_cases(self, xp): def check_nan(res): assert xp.isnan(res.statistic) assert xp.isnan(res.pvalue) message = 'One or more sample arguments is too small...' with eager_warns(SmallSampleWarning, match=message, xp=xp): res = stats.spearmanrho(xp.asarray([1]), xp.asarray([2])) check_nan(res) x = xp.asarray([1, 1, 1, 1, 1]) y = xp.asarray([1, 2, 3, 4, 5]) message = 'An input array is constant; the correlation coefficient...' with eager_warns(stats.ConstantInputWarning, match=message, xp=xp): res = stats.spearmanrho(x, y) check_nan(res) with eager_warns(stats.ConstantInputWarning, match=message, xp=xp): res = stats.spearmanrho(y, x) check_nan(res) class RobustSlopesTest: # 'iv, mask, warnings, consistency, gh19678, mstats' def test_input_validation(self, xp): pfun = getattr(stats, self.pfun) other_method = 'joint' if pfun == stats.theilslopes else 'hierarchical' msg = (f"method must be either '{other_method}' or 'separate'. " "'joint_separate' is invalid.") with pytest.raises(ValueError, match=msg): pfun(xp.asarray([0, 1, 1]), method='joint_separate') @skip_xp_invalid_arg @pytest.mark.parametrize("method", ['separate', 'other']) def test_mask(self, method): pfun = getattr(stats, self.pfun) if method == 'other': method = 'joint' if self.pfun =='theilslopes' else 'hierarchical' # Test for correct masking. mask = np.asarray([False, False, True, False]) y = np.ma.array([0, 1, 100, 1], mask=mask) res = pfun(y, method=method) ref = pfun(y[~mask], method=method) assert_allclose(res.slope, ref.slope) assert_allclose(res.intercept, ref.intercept) def test_degenerate(self, xp): # Test with degenerate input; see gh-15943 pfun = getattr(stats, self.pfun) res = pfun(xp.asarray([0, 1]), xp.asarray([0, 0])) assert xp.all(xp.isnan(xp.stack(res))) res = pfun(xp.asarray([0, 0, 0]), xp.asarray([0, 1, 0])) xp_assert_equal(res[0], xp.asarray(0.)) xp_assert_equal(res[1], xp.asarray(0.)) if pfun == stats.theilslopes: xp_assert_equal(res[2], xp.asarray(xp.nan)) xp_assert_equal(res[3], xp.asarray(xp.nan)) def test_namedtuple_consistency(self, xp): """ Simple test to ensure tuple backwards-compatibility of the returned object. """ pfun = getattr(stats, self.pfun) y = xp.asarray([1, 2, 4]) x = xp.asarray([4, 6, 8]) result = pfun(y, x) # note both returned values are distinct here xp_assert_equal(result[0], result.slope) xp_assert_equal(result[1], result.intercept) if pfun == stats.theilslopes: xp_assert_equal(result[2], result.low_slope) xp_assert_equal(result[3], result.high_slope) def test_gh19678_uint8(self, xp): # `theilslopes` returned unexpected results when `y` was an unsigned type. # Check that this is resolved. pfun = getattr(stats, self.pfun) rng = np.random.default_rng(2549824598234528) y = xp.asarray(rng.integers(0, 255, size=10, dtype=np.uint8)) res = pfun(y, y) xp_assert_close(res.slope, xp.asarray(1.)) @pytest.mark.parametrize("method", ['separate', 'other']) @pytest.mark.parametrize("case_number", [0, 1, 2, 3]) def test_against_mstats(self, method, case_number, xp): pfun = getattr(stats, self.pfun) if method == 'other': method = 'joint' if self.pfun =='theilslopes' else 'hierarchical' rng = np.random.default_rng(349824598234528554) match case_number: case 0: # no x x = None y = rng.random(100) case 1: # no ties x = rng.random(100) y = rng.random(100) case 2: # no x ties x = rng.random(100) y = rng.integers(50, size=100) case 3: # ties in x and y x = rng.integers(25, size=100) y = rng.integers(50, size=100) ref = pfun(y, x, method=method) y, x = xp.asarray(y), x if x is None else xp.asarray(x) if (is_array_api_strict(xp) and xp.isdtype(y.dtype, 'integral') and xp.isdtype(x.dtype, 'real floating')): # array API strict doesn't allow mixed type promotion, so promote for it y = xp.astype(y, xp.float64) dtype = xp_result_type(y, x, force_floating=True, xp=xp) res = pfun(y, x, method=method, axis=-1) assert res.slope != 0 # ensure randomly generated test case isn't trivial xp_assert_close(res.slope, xp.asarray(ref.slope, dtype=dtype)) xp_assert_close(res.intercept, xp.asarray(ref.intercept, dtype=dtype)) if pfun == stats.theilslopes: xp_assert_close(res.low_slope, xp.asarray(ref.low_slope, dtype=dtype)) xp_assert_close(res.high_slope, xp.asarray(ref.high_slope, dtype=dtype)) @make_xp_test_case(stats.theilslopes) class TestTheilslopes(RobustSlopesTest): pfun = 'theilslopes' def test_theilslopes(self, xp): # Test for basic slope and intercept. y = xp.asarray([0, 1, 1]) slope, intercept, lower, upper = stats.theilslopes(y) xp_assert_close(slope, xp.asarray(0.5)) xp_assert_close(intercept, xp.asarray(0.5)) slope, intercept, lower, upper = stats.theilslopes(y, method='joint') xp_assert_close(slope, xp.asarray(0.5)) xp_assert_close(intercept, xp.asarray(0.0)) # Test of confidence intervals from example in Sen (1968). x = xp.asarray([1, 2, 3, 4, 10, 12, 18]) y = xp.asarray([9, 15, 19, 20, 45, 55, 78]) slope, intercept, lower, upper = stats.theilslopes(y, x, 0.07) xp_assert_close(slope, xp.asarray(4.0)) xp_assert_close(intercept, xp.asarray(4.0)) xp_assert_close(upper, xp.asarray(4.38), rtol=5e-3) xp_assert_close(lower, xp.asarray(3.71), rtol=5e-3) slope, intercept, lower, upper = stats.theilslopes(y, x, 0.07, method='joint') xp_assert_close(slope, xp.asarray(4.0)) xp_assert_close(intercept, xp.asarray(6.0)) xp_assert_close(upper, xp.asarray(4.38), rtol=5e-3) xp_assert_close(lower, xp.asarray(3.71), rtol=5e-3) @make_xp_test_case(stats.siegelslopes) class TestSiegelslopes(RobustSlopesTest): pfun = 'siegelslopes' def test_siegelslopes(self, xp): # method should be exact for straight line y = 2 * xp.arange(10.) + 0.5 slope, intercept = stats.siegelslopes(y) xp_assert_close(slope, xp.asarray(2.0)) xp_assert_close(intercept, xp.asarray(0.5)) slope, intercept = stats.siegelslopes(y, method='separate') xp_assert_close(slope, xp.asarray(2.0)) xp_assert_close(intercept, xp.asarray(0.5)) x = 2 * xp.arange(10.) y = 5 * x - 3.0 slope, intercept = stats.siegelslopes(y, x) xp_assert_close(slope, xp.asarray(5.0)) xp_assert_close(intercept, xp.asarray(-3.0)) slope, intercept = stats.siegelslopes(y, x, method='separate') xp_assert_close(slope, xp.asarray(5.0)) xp_assert_close(intercept, xp.asarray(-3.0)) # method is robust to outliers: breakdown point of 50% y = xpx.at(y)[:4].set(1000.) xp_assert_close(slope, xp.asarray(5.0)) xp_assert_close(intercept, xp.asarray(-3.0)) # if there are no outliers, results should be comparable to linregress x = np.arange(10.) y = -2.3 + 0.3 * x + stats.norm.rvs(size=10, random_state=231) slope_ols, intercept_ols, _, _, _ = stats.linregress(x, y) y, x = xp.asarray(y), xp.asarray(x) slope, intercept = stats.siegelslopes(y, x) xp_assert_close(slope, xp.asarray(slope_ols), rtol=0.1) xp_assert_close(intercept, xp.asarray(intercept_ols), rtol=0.1) slope, intercept = stats.siegelslopes(y, x, method='separate') xp_assert_close(slope, xp.asarray(slope_ols), rtol=0.1) xp_assert_close(intercept, xp.asarray(intercept_ols), rtol=0.1) def test_method(self, xp): # Test that distinguishes between the methods. Reference values generated with # SciPy 1.17 (results unchanged since at least 1.12, so assumed "correct") to # detect *changes*. y = xp.asarray([2, 1, 3, 0, 4]) slope, intercept = stats.siegelslopes(y, method='separate') xp_assert_close(slope, xp.asarray(0.25)) xp_assert_close(intercept, xp.asarray(1.75)) slope, intercept = stats.siegelslopes(y, method='hierarchical') xp_assert_close(slope, xp.asarray(0.25)) # always the same between the methods xp_assert_close(intercept, xp.asarray(2.0))