Last active
June 13, 2019 00:23
-
-
Save WarrenWeckesser/26d7ef31e52561a695b5efad854f17bf to your computer and use it in GitHub Desktop.
Test utilities for numpy's hypergeometric distribution.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| # | |
| # Test utilities for numpy's hypergeometric distribution. | |
| # | |
| # Author: Warren Weckesser | |
| # License: BSD 2 clause | |
| # | |
| import mpmath | |
| import numpy as np | |
| from numpy.random import Generator, Xoshiro256 | |
| from mpsci.distributions import hypergeometric | |
| from mpsci.stats import gtest | |
| mpmath.mp.dps = 20 | |
| def make_tester(expected, minsize=100): | |
| n = len(expected) | |
| low = expected < minsize | |
| numlow = np.count_nonzero(low) | |
| total_low = expected[low].sum() | |
| if numlow == 0: | |
| newlen = n | |
| exp2 = expected.copy() | |
| def tester(observed): | |
| if len(observed) != n: | |
| raise ValueError("len(observed) must equal len(expected)") | |
| return numlow, total_low, newlen, gtest(observed, exp2) | |
| elif total_low >= minsize: | |
| # Aggregate all the low expected values into a | |
| # single bin. | |
| # This could be smarter. It joins all the low | |
| # bins into a single bin, even if the sum of the | |
| # expected values of the aggregated bins is more | |
| # than 2*minsize. | |
| newlen = n - numlow + 1 | |
| exp2 = np.empty(newlen, dtype=expected.dtype) | |
| exp2[:-1] = expected[~low] | |
| exp2[-1] = total_low | |
| def tester(observed): | |
| if len(observed) != n: | |
| raise ValueError("len(observed) must equal len(expected)") | |
| obs2 = np.empty(newlen, dtype=observed.dtype) | |
| obs2[:-1] = observed[~low] | |
| obs2[-1] = observed[low].sum() | |
| return numlow, total_low, newlen, gtest(obs2, exp2) | |
| else: | |
| # If here, the sum of all the low expected values | |
| # is still less than minsize. So we'll merge all | |
| # these into an existing bin. | |
| # XXX This just uses the last bin for which the expected | |
| # value is greater than minsize! | |
| newlen = n - numlow | |
| exp2 = np.empty(newlen, dtype=expected.dtype) | |
| exp2 = expected[~low] | |
| exp2[-1] += total_low | |
| def tester(observed): | |
| if len(observed) != n: | |
| raise ValueError("len(observed) must equal len(expected)") | |
| obs2 = np.empty(newlen, dtype=observed.dtype) | |
| obs2 = observed[~low] | |
| obs2[-1] += observed[low].sum() | |
| return numlow, total_low, newlen, gtest(obs2, exp2) | |
| return tester | |
| def compare(observed, expected, minsize=100): | |
| n = len(observed) | |
| if n != len(expected): | |
| raise ValueError("len(observed) must equal len(expected)") | |
| tester = make_tester(expected, minsize) | |
| return tester(observed) | |
| def hypergeometric_pvalues(good, bad, sample, size, ntrials, rng=None): | |
| if rng is None: | |
| gen = np.random.Generator(Xoshiro256()) | |
| else: | |
| gen = rng | |
| # The support is [mink, ..., maxk] (inclusive). | |
| mink = max(0, sample - bad) | |
| maxk = min(good, sample) | |
| expected = np.array([float(size*hypergeometric.pmf(k, good+bad, good, sample)) | |
| for k in range(mink, maxk+1)]) | |
| tester = make_tester(expected, minsize=100) | |
| pvals = [] | |
| for i in range(ntrials): | |
| s = gen.hypergeometric(good, bad, sample, size) | |
| observed = np.bincount(s, minlength=maxk + 1)[mink:] | |
| numlow, total_low, newlen, (stat, p) = tester(observed) | |
| pvals.append(float(p)) | |
| return pvals | |
| def hypergeometric_wide_range_test(size=100_000_000, rng=None): | |
| if rng is None: | |
| gen = np.random.Generator(Xoshiro256()) | |
| else: | |
| gen = rng | |
| print(" hypergeometric distribution test") | |
| print(" size =", size) | |
| print() | |
| print(" len total low g-test") | |
| print(" good bad nsample support nlow expected len p-value") | |
| for goodpar in [1e-1, 1e-2, 1e-3, 1e-4, 1e-5, 1e-6]: | |
| for badpar in [1e-1, 1e-2, 1e-3, 1e-4, 1e-5, 1e-6]: | |
| for k in range(2): | |
| # Generate a random set of parameters (good, bad, sample) | |
| # for the hypergeometric distribution. | |
| good = gen.geometric(goodpar) | |
| bad = gen.geometric(badpar) | |
| while True: | |
| frac = gen.beta(0.75, 2.5) | |
| sample = int(frac*(good + bad) + 0.5) | |
| if sample > 0: | |
| break | |
| # The support is [mink, ..., maxk] (inclusive). | |
| mink = max(0, sample - bad) | |
| maxk = min(good, sample) | |
| s = gen.hypergeometric(good, bad, sample, size) | |
| h = np.bincount(s, minlength=maxk + 1) | |
| expected = np.array([float(size*hypergeometric.pmf(k, good+bad, good, sample)) | |
| for k in range(mink, maxk+1)]) | |
| numlow, total_low, gtestlen, (stat, p) = compare(h[mink:], expected) | |
| print("%10d %10d %10d %8d" % (good, bad, sample, maxk - mink + 1), end='') | |
| print(" %8d %10.3f" % (numlow, total_low), end=' ') | |
| print(" %5d " % gtestlen, end='') | |
| print(" %8.6f" % p) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment