def test_confint_2indep_propcis():
# unit tests compared to R package PropCis
# alpha = 0.05
count1, nobs1 = 7, 34
count2, nobs2 = 1, 34
# > library(PropCIs)
# > diffscoreci(7, 34, 1, 34, 0.95)
ci = 0.0270416, 0.3452912
ci1 = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare="diff",
method="score",
correction=False)
assert_allclose(ci1, ci, atol=0.002) # lower agreement (iterative)
# > wald2ci(7, 34, 1, 34, 0.95, adjust="AC")
ci = 0.01161167, 0.32172166
ci1 = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare="diff",
method="agresti-caffo")
assert_allclose(ci1, ci, atol=6e-7)
# > wald2ci(7, 34, 1, 34, 0.95, adjust="Wald")
ci = 0.02916942, 0.32377176
ci1 = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare="diff",
method="wald",
correction=False)
assert_allclose(ci1, ci, atol=6e-7)
# > orscoreci(7, 34, 1, 34, 0.95)
ci = 1.246309, 56.486130
ci1 = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare="odds-ratio",
method="score",
correction=True)
assert_allclose(ci1, ci, rtol=5e-4) # lower agreement (iterative)
# > riskscoreci(7, 34, 1, 34, 0.95)
ci = 1.220853, 42.575718
ci1 = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare="ratio",
method="score",
correction=False)
assert_allclose(ci1, ci, atol=6e-7)
def test_test_2indep():
# this checks the pvalue of the hypothesis test at value equal to the
# confidence limit
alpha = 0.05
count1, nobs1 = 7, 34
count2, nobs2 = 1, 34
methods_both = [
('diff', 'agresti-caffo'),
# ('diff', 'newcomb'), # only confint
('diff', 'score'),
('diff', 'wald'),
('ratio', 'log'),
('ratio', 'log-adjusted'),
('ratio', 'score'),
('odds-ratio', 'logit'),
('odds-ratio', 'logit-adjusted'),
('odds-ratio', 'logit-smoothed'),
('odds-ratio', 'score'),
]
for co, method in methods_both:
low, upp = confint_proportions_2indep(count1, nobs1, count2, nobs2,
compare=co, method=method,
alpha=alpha, correction=False)
res = smprop.test_proportions_2indep(
count1, nobs1, count2, nobs2, value=low, compare=co,
method=method, correction=False)
assert_allclose(res.pvalue, alpha, atol=1e-10)
res = smprop.test_proportions_2indep(
count1, nobs1, count2, nobs2, value=upp, compare=co,
method=method, correction=False)
assert_allclose(res.pvalue, alpha, atol=1e-10)
_, pv = smprop.test_proportions_2indep(
count1, nobs1, count2, nobs2, value=upp, compare=co,
method=method, alternative='smaller',
correction=False, return_results=False)
assert_allclose(pv, alpha / 2, atol=1e-10)
_, pv = smprop.test_proportions_2indep(
count1, nobs1, count2, nobs2, value=low, compare=co,
method=method, alternative='larger',
correction=False, return_results=False)
assert_allclose(pv, alpha / 2, atol=1e-10)
# test Miettinen/Nurminen small sample correction
co, method = 'ratio', 'score'
low, upp = confint_proportions_2indep(count1, nobs1, count2, nobs2,
compare=co, method=method,
alpha=alpha, correction=True)
res = smprop.test_proportions_2indep(
count1, nobs1, count2, nobs2, value=low, compare=co,
method=method, correction=True)
assert_allclose(res.pvalue, alpha, atol=1e-10)
def test_equivalence_2indep():
# this checks the pvalue of the equivalence test at value equal to the
# confidence limit
alpha = 0.05
count1, nobs1 = 7, 34
count2, nobs2 = 1, 34
methods_both = [
('diff', 'agresti-caffo'),
# ('diff', 'newcomb'), # only confint
('diff', 'score'),
('diff', 'wald'),
('ratio', 'log'),
('ratio', 'log-adjusted'),
('ratio', 'score'),
('odds-ratio', 'logit'),
('odds-ratio', 'logit-adjusted'),
('odds-ratio', 'logit-smoothed'),
('odds-ratio', 'score'),
]
for co, method in methods_both:
low, upp = confint_proportions_2indep(count1, nobs1, count2, nobs2,
compare=co, method=method,
alpha=2 * alpha,
correction=False)
res = smprop.tost_proportions_2indep(
count1, nobs1, count2, nobs2, low, upp, compare=co,
method=method, correction=False)
assert_allclose(res.pvalue, alpha, atol=1e-10)
def _ci(self, row, alpha_column: str) -> Tuple[float, float]:
return confint_proportions_2indep(count1=row[self._numerator + SFX2],
nobs1=row[self._denominator + SFX2],
count2=row[self._numerator + SFX1],
nobs2=row[self._denominator + SFX1],
alpha=row[alpha_column],
compare='diff',
method='wald')
def two_props_diff_conf_interval(values1: np.ndarray, values2: np.ndarray,
conf_level: float) -> tuple:
"""Calculates the confidence interval for the diff between two proportions
Args:
values1 (np.array): sample 1 binary(0/1) values
values2 (np.array): sample 2 binary(0/1) values
conf_level (float): confidence level
Returns:
tuple: lower and upper values of confidence interval
"""
count1, count2 = values1.sum(), values2.sum()
nobs1, nobs2 = len(values1), len(values2)
ci = prop_stats.confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
alpha=1 - conf_level,
method="wald")
return ci
def test_confint_2indep():
# alpha = 0.05
count1, nobs1 = 7, 34
count2, nobs2 = 1, 34
# result tables from Fagerland et al 2015
'''
diff:
Wald 0.029 0.32 0.29
Agresti–Caffo 0.012 0.32 0.31
Newcombe hybrid score 0.019 0.34 0.32
Miettinen–Nurminen asymptotic score 0.028 0.34 0.31
Santner–Snell exact unconditional -0.069 0.41 0.48
Chan–Zhang exact unconditional 0.019 0.36 0.34
Agresti–Min exact unconditional 0.024 0.35 0.33
ratio:
Katz log 0.91 54 4.08
Adjusted log 0.92 27 3.38
Inverse sinh 1.17 42 3.58
Koopman asymptotic score 1.21 43 3.57
Chan–Zhang 1.22 181 5.00
Agresti–Min 1.15 89 4.35
odds-ratio
Woolf logit 0.99 74 4.31
Gart adjusted logit 0.98 38 3.65
Independence-smoothed logit 0.99 60 4.11
Cornfield exact conditional 0.97 397 6.01
Cornfield mid-p 1.19 200 5.12
Baptista–Pike exact conditional 1.00 195 5.28
Baptista–Pike mid-p 1.33 99 4.31
Agresti–Min exact unconditional 1.19 72 4.10
''' # pylint: disable=W0105
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
method='newcomb',
compare='diff',
alpha=0.05)
# one decimal to upp added from regression result
assert_allclose(ci, [0.019, 0.340], atol=0.005)
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
method='wald',
compare='diff',
alpha=0.05)
assert_allclose(ci, [0.029, 0.324], atol=0.005)
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
method='agresti-caffo',
compare='diff',
alpha=0.05)
assert_allclose(ci, [0.012, 0.322], atol=0.005)
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare='diff',
method='score',
correction=True)
assert_allclose(ci, [0.028, 0.343], rtol=0.03)
# ratio
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare='ratio',
method='log')
assert_allclose(ci, [0.91, 54], rtol=0.01)
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare='ratio',
method='log-adjusted')
assert_allclose(ci, [0.92, 27], rtol=0.01)
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare='ratio',
method='score',
correction=False)
assert_allclose(ci, [1.21, 43], rtol=0.01)
# odds-ratio
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare='or',
method='logit')
assert_allclose(ci, [0.99, 74], rtol=0.01)
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare='or',
method='logit-adjusted')
assert_allclose(ci, [0.98, 38], rtol=0.01)
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare='or',
method='logit-smoothed')
assert_allclose(ci, [0.99, 60], rtol=0.01)
ci = confint_proportions_2indep(count1,
nobs1,
count2,
nobs2,
compare='odds-ratio',
method='score',
correction=True)
# regression test
assert_allclose(ci, [1.246622, 56.461576], rtol=0.01)
