def test_weightstats_ddof_tests(self):
# explicit test that ttest and confint are independent of ddof
# one sample case
x1_2d = self.x1_2d
w1 = self.w1
d1w_d0 = DescrStatsW(x1_2d, weights=w1, ddof=0)
d1w_d1 = DescrStatsW(x1_2d, weights=w1, ddof=1)
d1w_d2 = DescrStatsW(x1_2d, weights=w1, ddof=2)
# check confint independent of user ddof
res0 = d1w_d0.ttest_mean()
res1 = d1w_d1.ttest_mean()
res2 = d1w_d2.ttest_mean()
# concatenate into one array with np.r_
assert_almost_equal(np.r_[res1], np.r_[res0], 14)
assert_almost_equal(np.r_[res2], np.r_[res0], 14)
res0 = d1w_d0.ttest_mean(0.5)
res1 = d1w_d1.ttest_mean(0.5)
res2 = d1w_d2.ttest_mean(0.5)
assert_almost_equal(np.r_[res1], np.r_[res0], 14)
assert_almost_equal(np.r_[res2], np.r_[res0], 14)
# check confint independent of user ddof
res0 = d1w_d0.tconfint_mean()
res1 = d1w_d1.tconfint_mean()
res2 = d1w_d2.tconfint_mean()
assert_almost_equal(res1, res0, 14)
assert_almost_equal(res2, res0, 14)
def test_weightstats_ddof_tests(self):
# explicit test that ttest and confint are independent of ddof
# one sample case
x1_2d = self.x1_2d
w1 = self.w1
d1w_d0 = DescrStatsW(x1_2d, weights=w1, ddof=0)
d1w_d1 = DescrStatsW(x1_2d, weights=w1, ddof=1)
d1w_d2 = DescrStatsW(x1_2d, weights=w1, ddof=2)
#check confint independent of user ddof
res0 = d1w_d0.ttest_mean()
res1 = d1w_d1.ttest_mean()
res2 = d1w_d2.ttest_mean()
# concatenate into one array with np.r_
assert_almost_equal(np.r_[res1], np.r_[res0], 14)
assert_almost_equal(np.r_[res2], np.r_[res0], 14)
res0 = d1w_d0.ttest_mean(0.5)
res1 = d1w_d1.ttest_mean(0.5)
res2 = d1w_d2.ttest_mean(0.5)
assert_almost_equal(np.r_[res1], np.r_[res0], 14)
assert_almost_equal(np.r_[res2], np.r_[res0], 14)
#check confint independent of user ddof
res0 = d1w_d0.tconfint_mean()
res1 = d1w_d1.tconfint_mean()
res2 = d1w_d2.tconfint_mean()
assert_almost_equal(res1, res0, 14)
assert_almost_equal(res2, res0, 14)
def summarize(name, test=False, space_creator=space_optimized):
print("---------------------")
print(f"{name}> summarizing")
files = list(glob.glob(f"output/{name}/*_results.txt"))
print(f"{name}> {len(files)} files found")
failed = 0
timed_out = 0
memory_error = 0
results = []
for file in files:
log_file = file.replace("_results.txt", "_log.txt")
with open(log_file) as f:
lines = "\n".join(f.readlines())
if "out of memory" in lines:
failed += 1
memory_error += 1
print(f"{name}> {log_file} out of memory")
continue
with open(file) as f:
lines = f.readlines()
time_taken = int(lines[3]) / 1000 / 60 # minutes
if lines[5].startswith("False"):
print(f"{name}> {file} timed out")
failed += 1
timed_out += 1
continue
results.append(time_taken)
if test:
check(log_file, stop_on_error=True, space_creator=space_creator)
success_count = len(results)
print(f"{name}> {success_count} successes")
if success_count != 0:
d = DescrStatsW(results)
print(f"{name}> mean={d.mean}, std_mean={d.std_mean}")
print(f"{name}> confidence interval", d.tconfint_mean())
print(
f"{name}> {failed} failed, {failed / (success_count + failed) * 100}%")
print(f"{name}> {timed_out} timed out")
print(f"{name}> {memory_error} memory error")
print(
f"{name}> binomial success ci={proportion_confint(success_count, success_count + failed, method='wilson')}"
)
frame = frame_of(results, name)
plt.figure()
sns.boxplot(data=frame, y="time")
plt.savefig(f"output/fig{name}.png")
plt.figure()
sns.displot(frame["time"])
plt.ylim(0, 350)
plt.savefig(f"output/fig{name}-dist.png")
return results
def grouped_weights_statsdf(df, statscols, groupbycol, weightscol):
"""generates df with weighted means and 95% CI by groupbycol for cols in statscols
Parameters
----------
df : df
df to be weigthed
statscols : list
cols/outcomes for weigthed stats
groupbycol : str
column name in df that defines groups
weightscol : str
column name in df with weigths
Returns
-------
df
multi-indexed df with outcome and groups as index
stats generated: weighted mean, upper bound (95 CI), lower bound (95% CI), weighted n by group, total n unweighted
"""
alldata = pd.DataFrame()
for c in statscols:
cdf = df.dropna(subset=[c])
nrobs = len(cdf)
grouped = cdf.groupby(groupbycol)
stats = {}
means = []
lower = []
upper = []
nrobs_gr = []
groups = list(grouped.groups.keys())
for gr in groups:
stats = DescrStatsW(grouped.get_group(gr)[c],
weights=grouped.get_group(gr)[weightscol],
ddof=0)
means.append(stats.mean)
lower.append(stats.tconfint_mean()[0])
upper.append(stats.tconfint_mean()[1])
nrobs_gr.append(stats.nobs)
weightedstats = pd.DataFrame([means, lower, upper, nrobs_gr],
columns=groups,
index=[
'weighted mean', 'lower bound',
'upper bound', 'wei_n__group'
]).T
weightedstats['tot_n_unweigthed'] = nrobs
weightedstats['outcome'] = c
weightedstats.index.name = 'groups'
colstats = weightedstats.reset_index()
colstats = colstats.set_index(['outcome', 'groups'])
alldata = pd.concat([alldata, colstats])
return alldata
def one_t_test(pdf, data_measlevs, var_name, test_value=0):
"""One sample t-test
arguments:
var_name (str):
Name of the variable to test.
test_value (numeric):
Test against this value.
return:
text_result (html str):
Result in APA format.
image (matplotlib):
Bar chart with mean and confidence interval.
"""
text_result = ''
data = pdf[var_name].dropna()
if data_measlevs[var_name] in ['int', 'unk']:
if data_measlevs[var_name] == 'unk':
text_result += warn_unknown_variable
if len(set(data)) == 1:
return _('One sample t-test cannot be run for constant variable.\n'
), None
data = pdf[var_name].dropna()
descr = DescrStatsW(data)
t, p, df = descr.ttest_mean(float(test_value))
if LooseVersion(csc.versions['statsmodels']) >= LooseVersion('0.5'):
# Or we could use confidence_interval_t
cil, cih = descr.tconfint_mean()
ci = (cih - cil) / 2
prec = cs_util.precision(data) + 1
ci_text = '[%0.*f, %0.*f]' % (prec, cil, prec, cih)
else:
ci = 0 # only with statsmodels
ci_text = _(
'Sorry, newer statsmodels module is required for confidence interval.\n'
)
text_result += _('One sample t-test against %g') % float(
test_value) + ': <i>t</i>(%d) = %0.3g, %s\n' % (df, t,
cs_util.print_p(p))
# Graph
image = cs_chart.create_variable_population_chart(data, var_name, ci)
else:
text_result += _(
'One sample t-test is computed only for interval variables.')
image = None
return ci_text, text_result, image
def grouped_weights_statscol (df, statscol, groupbycol, weightscol):
df.dropna(subset=[statscol], inplace=True)
nrobs=len(df)
grouped=df.groupby(groupbycol)
stats={}
means=[]
lower=[]
upper=[]
groups=list(grouped.groups.keys())
for gr in groups:
stats=DescrStatsW(grouped.get_group(gr)[statscol], weights=grouped.get_group(gr)[weightscol], ddof=0)
means.append(stats.mean)
lower.append(stats.tconfint_mean()[0])
upper.append(stats.tconfint_mean()[1])
weightedstats=pd.DataFrame([means, lower, upper], columns=groups, index=['weighted mean', 'lower bound', 'upper bound']).T
weightedstats['numberofobs']=nrobs
return weightedstats
def confidence_interval_t(data, ci_only=True):
"""95%, two-sided CI based on t-distribution
http://statsmodels.sourceforge.net/stable/_modules/statsmodels/stats/weightstats.html#DescrStatsW.tconfint_mean
"""
# FIXME is this solution slow? Should we write our own CI function?
if LooseVersion(csc.versions['statsmodels']) >= LooseVersion('0.5'):
descr = DescrStatsW(data)
cil, cih = descr.tconfint_mean()
ci = (cih - cil) / 2
else:
cil = cih = ci = [None
for i in data] # FIXME maybe this one is not correct
if ci_only:
if isinstance(data, pd.Series):
return ci # FIXME this one is for series? The other is for dataframes?
elif isinstance(data, pd.DataFrame):
return pd.Series(ci, index=data.columns)
# without var names the call from comp_group_graph_cum fails
else:
return ci, cil, cih
my_knn_socres = cross_val_score(KNeighborsRegressor(n_neighbors=5),
X,
y,
cv=LeaveOneOut(),
scoring='neg_mean_squared_error')
(-my_lm_scores.mean())**0.5
#> 15.697306009399101 # 線形回帰分析
(-my_knn_socres.mean())**0.5
#> 16.07308308943869 # K最近傍法
my_df = pd.DataFrame({'lm': -my_lm_scores, 'knn': -my_knn_socres})
my_df.head()
#> lm knn
#> 0 18.913720 108.16
#> 1 179.215044 0.64
#> 2 41.034336 64.00
#> 3 168.490212 184.96
#> 4 5.085308 0.00
my_df.boxplot().set_ylabel("$r^2$")
from statsmodels.stats.weightstats import DescrStatsW
d = DescrStatsW(my_df.lm - my_df.knn)
d.ttest_mean()[1] # p値
#> 0.6952755720536115
d.tconfint_mean(alpha=0.05, alternative='two-sided') # 信頼区間
#> (-72.8275283312228, 48.95036023665703)
def __call__(self, alpha, df):
stats = DescrStatsW(df)
(lower, upper) = stats.tconfint_mean(alpha=alpha)
return Band(lower, upper)
np.random.seed(75243)
temp = nota_media_dos_filmes_com_pelo_menos_10_votos.sample(frac=1)
medias = [temp[0:i].mean() for i in range(1, len(temp))]
plt.plot(medias)
from statsmodels.stats.weightstats import zconfint
zconfint(nota_media_dos_filmes_com_pelo_menos_10_votos)
from statsmodels.stats.weightstats import DescrStatsW
descr_todos_com_10_votos = DescrStatsW(nota_media_dos_filmes_com_pelo_menos_10_votos)
descr_todos_com_10_votos.tconfint_mean()
"""# Vamos ver o filme 1..."""
filmes = pd.read_csv("movies.csv")
filmes.query("movieId==1")
notas1 = notas.query("movieId == 1")
notas1.head()
ax = sns.distplot(notas1.rating)
ax.set(xlabel="Nota", ylabel="Densidade")
ax.set_title("Distribuição das notas para o Toy Story")
ax = sns.boxplot(notas1.rating)
ax.set(xlabel="Nota")
]
Y = [
35.4, 34.6, 31.1, 32.4, 33.3, 34.7, 35.3, 34.3, 32.1, 28.3, 33.3, 30.5,
32.6, 33.3, 32.2
]
a = 0.05 # 有意水準（デフォルト） = 1 - 信頼係数
alt = 'two-sided' # 両側検定（デフォルト）
# 左片側検定なら'smaller'
# 右片側検定なら'larger'
d = DescrStatsW(np.array(X) - np.array(Y)) # 対標本の場合
d.ttest_mean(alternative=alt)[1] # p値
#> 0.0006415571512322235
d.tconfint_mean(alpha=a, alternative=alt) # 信頼区間
#> (-3.9955246743198867, -1.3644753256801117)
c = CompareMeans(DescrStatsW(X), DescrStatsW(Y)) # 対標本でない場合
ve = 'pooled' # 等分散を仮定する（デフォルト）．仮定しないなら'unequal'．
c.ttest_ind(alternative=alt, usevar=ve)[1] # p値
#> 0.000978530937238609
c.tconfint_diff(alpha=a, alternative=alt, usevar=ve) # 信頼区間
#> (-4.170905570517185, -1.1890944294828283)
### 4.4.4 独立性の検定（カイ2乗検定）
import pandas as pd
my_url = ('https://raw.githubusercontent.com/taroyabuki'
diagnostico_b = df.query("diagnosis == 'B'")
# Efetuando o Zteste para a média (Comparando os resultados)
ztest(diagnostico_m['mean_radius'], value = diagnostico_m['mean_radius'].mean())
ztest(diagnostico_m['mean_radius'], value = diagnostico_b['mean_radius'].mean())
# Gerando o intervalo de confiança
zconfint(diagnostico_m['mean_radius'])
zconfint(diagnostico_b['mean_radius'])
"""----------------------------------------------------------------------------
T Test
"""
diagnostico_m = df.query("diagnosis == 'M'")
diagnostico_b = df.query("diagnosis == 'B'")
# Aplicando o teste
resultados_m = DescrStatsW(diagnostico_m['mean_radius'])
resultados_b = DescrStatsW(diagnostico_b['mean_radius'])
# Gerando o intervalo de confiança
resultados_m.tconfint_mean()
resultados_b.tconfint_mean()
