fast-fisher 0.0.4

Credit

This is heavily based on the work of painyeph.

I basically added numba compilation and easier installation. Nevertheless, it makes a big difference speed-wise!

Installation

pip install git+https://github.com/MrTomRod/fast-fisher

# or

pip install fast-fisher  # from https://pypi.org/project/fast-fisher/

Usage

Recommended usage:

from fast_fisher import fast_fisher_exact, odds_ratio

a, b, c, d = 15, 32, 25, 46

pvalue = fast_fisher_exact(a, b, c, d, alternative='two-sided')
odds = odds_ratio(a, b, c, d)

Alternative usages:

fast_fisher_exact_compatibility has the same syntax as scipy.stats.fisher_exact.

from math import isclose
from scipy.stats import fisher_exact
from fast_fisher import fast_fisher_exact_compatibility

table = [[15, 32], [25, 46]]

for alternative in ['two-sided', 'less', 'greater']:
    odds_s, pval_s = fisher_exact(table, alternative)
    odds_f, pval_f = fast_fisher_exact_compatibility(table, alternative)
    assert odds_s == odds_f
    assert isclose(pval_s, pval_f)  # not always true!

Advanced Usage

Speed

Comparison of

• scipy.stats.fisher_exact (scipy docs)
• fast_fisher.fast_fisher_python (this library)
• fast_fisher.fast_fisher_compiled (this library, compiled using numba)
• fast_fisher.fast_fisher_cpython (this library, compiled using cython)
• fisher.pvalue (from brentp/fishers_exact_test)

See benchmark.py.

While numba seems be marginally faster than the cython, cython is a much better build dependency. Therefore, by default, only the cython version is installed.

Precision

from numpy import log10, isinf
from scipy.stats import fisher_exact
from fast_fisher import fast_fisher

scipy_fisher = lambda t: fisher_exact([[t[0], t[1]], [t[2], t[3]]])[1]

print(f"{'contingency table':<30} {'fast -log10(pvalue)':<21} {'scipy -log10(pvalue)'}")
for exponent in range(0, 16):
    table = (100, 1, 10, 10 ** exponent)
    fast_mlog = fast_fisher.mlog10Test1t(*table)
    scipy_mlog = -log10(scipy_fisher(table))

    if isinf(scipy_mlog):
        scipy_mlog = 'failed to compute'

    print(f"{str(table):<30} {fast_mlog:<21} {scipy_mlog}")

contingency table              fast -log10(pvalue)   scipy -log10(pvalue)
(100, 1, 10, 1)                0.7268124553699258    0.7268124553698625
(100, 1, 10, 10)               7.831294376070296     7.831294376070258
(100, 1, 10, 100)              46.49556750272154     46.4955675027216
(100, 1, 10, 1000)             128.93472935802276    128.93472935802373
(100, 1, 10, 10000)            226.62104816785       226.62104816785057
(100, 1, 10, 100000)           326.3812661048001     failed to compute
(100, 1, 10, 1000000)          426.35719912501844    failed to compute
(100, 1, 10, 10000000)         526.3547915160074     failed to compute
(100, 1, 10, 100000000)        626.3545507841483     failed to compute
(100, 1, 10, 1000000000)       726.3545273226812     failed to compute
(100, 1, 10, 10000000000)      826.3545146294285     failed to compute
(100, 1, 10, 100000000000)     926.354158998833      failed to compute
(100, 1, 10, 1000000000000)    1026.3583095975994    failed to compute
(100, 1, 10, 10000000000000)   1126.3427388160835    failed to compute
(100, 1, 10, 100000000000000)  1226.447616894783     failed to compute
<input>:11: RuntimeWarning: divide by zero encountered in log10
Traceback (most recent call last):
  File "/usr/lib64/python3.10/code.py", line 90, in runcode
    exec(code, self.locals)
  File "<input>", line 10, in <module>
  File "fast_fisher/fast_fisher_cython.pyx", line 236, in fast_fisher.fast_fisher_cython.mlog10Test1t
    cpdef inline double mlog10Test1t(long long a, long long b, long long c, long long d) except *:
  File "fast_fisher/fast_fisher_cython.pyx", line 237, in fast_fisher.fast_fisher_cython.mlog10Test1t
    return mlnTest2t(a, a + b, a + c, a + b + c + d) / LN10
  File "fast_fisher/fast_fisher_cython.pyx", line 195, in fast_fisher.fast_fisher_cython.mlnTest2t
    raise OverflowError('the grand total of contingency table is too large')
OverflowError: the grand total of contingency table is too large