def compute(cls, observation, prediction):
"""Compute a t statistic and a p_value from an observation and a prediction."""
value, p_val = st.ttest_rel(prediction, observation)
power = pw.TTestPower().power(effect_size=value /
len(observation)**0.5,
nobs=len(observation),
alpha=0.05)
diffmean = numpy.mean(prediction) - numpy.mean(observation)
return StudentsPairedTestScore(value,
related_data={
"dof": len(observation) - 1,
"p_value": p_val,
"power": power,
"diffmean": diffmean
})
chi2_pow_R = 0.675077657003721
print('chi2_pow', chi2_pow, chi2_pow - chi2_pow_R)
chi2_pow = smp.GofChisquarePower().power(0.01, 100, 4, 0.05)
chi2_pow_R = 0.0505845519208533
print('chi2_pow', chi2_pow, chi2_pow - chi2_pow_R)
chi2_pow = smp.GofChisquarePower().power(2, 100, 4, 0.05)
chi2_pow_R = 1
print('chi2_pow', chi2_pow, chi2_pow - chi2_pow_R)
chi2_pow = smp.GofChisquarePower().power(0.9, 100, 4, 0.05)
chi2_pow_R = 0.999999999919477
print('chi2_pow', chi2_pow, chi2_pow - chi2_pow_R, 'lower precision ?')
chi2_pow = smp.GofChisquarePower().power(0.8, 100, 4, 0.05)
chi2_pow_R = 0.999999968205591
print('chi2_pow', chi2_pow, chi2_pow - chi2_pow_R)
def cohen_es(*args, **kwds):
print(
"You better check what's a meaningful effect size for your question.")
#BUG: after fixing 2.sided option, 2 rejection areas
tt_pow = smp.TTestPower().power(effect_size=0.01, nobs=nobs, alpha=0.05)
tt_pow_R = 0.05089485285965
# value from> pwr.t.test(d=0.01,n=80,sig.level=0.05,type="one.sample",alternative="two.sided")
print('tt_pow', tt_pow, tt_pow - tt_pow_R)
def calculate_paired_ttest_php(es, n_per_group):
power = smp.TTestPower().power(effect_size=es, alpha=0.05, nobs=n_per_group,
df=2*n_per_group-1, alternative='two-sided')
print("Power = {}".format(round(power, 3)))
def compute(cls, observation, prediction):
"""Compute a t statistic and a p_value from an observation and a prediction."""
p_mean = prediction['mean']
p_std = prediction['std']
p_n = prediction['n']
p_var = p_std**2
#2 samples t-test
if isinstance(observation, dict):
o_mean = observation['mean']
o_std = observation['std']
o_n = observation['n']
o_var = o_std**2
#If the 2 variances are too different, perform a Welch t-test
if p_var / o_var > 2 or o_var / p_var > 2:
value, p_val = st.ttest_ind_from_stats(p_mean,
p_std,
p_n,
o_mean,
o_std,
o_n,
equal_var=False)
vnp = p_var / p_n
vno = o_var / o_n
#Welch-Satherwaite equation to compute the degrees of freedom
dof = (vnp + vno)**2 / (vnp**2 / (p_n - 1) + vno**2 /
(o_n - 1))
#If the 2 variances are similar, perform a 2 sample independant Student t-test
else:
value, p_val = st.ttest_ind_from_stats(p_mean,
p_std,
p_n,
o_mean,
o_std,
o_n,
equal_var=True)
dof = o_n + p_n - 2
#Compute the statistical power of the test
power = pw.TTestIndPower().power(effect_size=CohenDScore.compute(
observation, prediction).score,
nobs1=p_n,
ratio=float(o_n) / p_n,
alpha=0.05)
#1 sample t-test
else:
value, p_val = st.ttest_ind_from_stats(p_mean,
p_std,
p_n,
observation,
std2=0,
nobs2=2,
equal_var=False)
#Compute the statistical power of the test
power = pw.TTestPower().power(effect_size=CohenDScore.compute(
{
"mean": observation,
"std": 0
}, prediction).score,
nobs=p_n,
alpha=0.05)
o_mean = observation
dof = p_n - 1
return StudentsTestScore(value,
related_data={
"dof": dof,
"p_value": p_val,
"power": power,
"diffmean": p_mean - o_mean
})