def test_result_attributes(self):
x = [1, 3, 5, 7, 9]
y = [2, 4, 6, 8, 10]
res = mstats.kruskal(x, y)
attributes = ('statistic', 'pvalue')
check_named_results(res, attributes, ma=True)
def test_result_attributes(self):
x = [1, 3, 5, 7, 9]
y = [2, 4, 6, 8, 10]
res = mstats.kruskal(x, y)
attributes = ('statistic', 'pvalue')
check_named_results(res, attributes, ma=True)
expenses_group1[val[0]] = int(val[1])
for val in total_spend_per_product_group2:
if val[0] in expenses_group2.keys():
expenses_group2[val[0]] += int(val[1])
else:
expenses_group2[val[0]] = int(val[1])
#print(expenses_group1, expenses_group2)
# Statistical study
# Kruskal-Wallis: Check whether the null hypothesis: "The average expenditure for each group is the same" is true
from scipy.stats.mstats import kruskal
from scipy.stats.mstats import mannwhitneyu
st, pvalue = kruskal([x[1] for x in customers_group1],
[x[1] for x in customers_group2
]) # x[1] contains the expenditure of each product
if pvalue < 0.05:
print(
"The null hypothesis: \n\t'The average expenditure for each group is the same'\nis False"
)
# Mann-Whitney for each pair of groups. Eg.: milk_group1 - milk_group2, delicatessen_group1 - delicatessen_group2
for product in range(2, 8):
product_group1 = []
product_group2 = []
# Get the expenditures per type of product
for value in customers_group1:
product_index = data[value[0]].index(value[1])
if product_index == product:
product_group1.append(value[1])
for value in customers_group2:
print("\n")
print(mannwhitneyu(all_data["Gaussian"], all_data["Finetuned\nFiji U-Net"])[1])
print(mannwhitneyu(all_data["Hessian"], all_data["Finetuned\nFiji U-Net"])[1])
print(
mannwhitneyu(all_data["Laplacian"], all_data["Finetuned\nFiji U-Net"])[1])
print(mannwhitneyu(all_data["Ilastik"], all_data["Finetuned\nFiji U-Net"])[1])
#print(np.median(all_data["Hessian"]), np.median(all_data["Ilastik"]), np.median(all_data["MitoSegNet"]))
#print(np.average(all_data["Hessian"]), np.average(all_data["Ilastik"]), np.average(all_data["MitoSegNet"]))
print(
kruskal(all_data["Gaussian"].tolist(), all_data["Hessian"].tolist(),
all_data["Laplacian"].tolist(), all_data["Ilastik"].tolist(),
all_data["MitoSegNet"].tolist(),
all_data["Finetuned\nFiji U-Net"].tolist()))
dt = posthoc_dunn([
all_data["Gaussian"].tolist(), all_data["Hessian"].tolist(),
all_data["Laplacian"].tolist(), all_data["Ilastik"].tolist(),
all_data["MitoSegNet"].tolist(),
all_data["Finetuned\nFiji U-Net"].tolist()
])
dt.to_excel("ed_posthoc.xlsx")
# pooled standard deviation for calculation of effect size (cohen's d)
def cohens_d(data1, data2):
print(mannwhitneyu(m_l, ga_l)[1])
print(mannwhitneyu(m_l, il_l)[1])
print(mannwhitneyu(m_l, u_l_pt)[1])
#print(mannwhitneyu(m_l, u_l)[1])
"""
all_data["Gaussian"] = ga_l
all_data["Hessian"] = h_l
all_data["Laplacian"] = la_l
all_data["Ilastik"] = il_l
all_data["MitoSegNet"] = m_l
all_data["Finetuned\nFiji U-Net"] = u_l_pt
"""
p_val = kruskal(ga_l, h_l, la_l, il_l, m_l, u_l_pt)
print(p_val)
dt = posthoc_dunn([ga_l, h_l, la_l, il_l, m_l, u_l_pt])
#dt = posthoc_dunn(all_data, val_col="MitoSegNet", group_col="Ilastik")
print(dt)
dt.to_excel("dc_posthoc.xlsx")
print("\n")
print(mannwhitneyu(u_l_pt, h_l)[1])
print(mannwhitneyu(u_l_pt, la_l)[1])
print(mannwhitneyu(u_l_pt, ga_l)[1])
print(mannwhitneyu(u_l_pt, il_l)[1])
print("\n")
dice_l.append(dice)
all_data[folder] = dice_l
#print(all_data)
#all_data.to_csv(path + "/dice_coefficient_table.csv")
"""
# checking if data is normally distributed
for seg in all_data:
print(seg, normaltest(all_data[seg]))
"""
# hypothesis testing
kwt = kruskal([
all_data["MitoSegNet"], all_data["Finetuned Fiji U-Net"],
all_data["Ilastik"], all_data["Gaussian"], all_data["Hessian"],
all_data["Laplacian"]
])
#print(kwt)
dt = posthoc_dunn([
all_data["MitoSegNet"], all_data["Finetuned Fiji U-Net"],
all_data["Ilastik"], all_data["Gaussian"], all_data["Hessian"],
all_data["Laplacian"]
])
#print(dt)
#dt.to_excel("dc_posthoc.xlsx")
#print(cohens_d(all_data["MitoSegNet"], all_data["Ilastik"]))
# pos_y and pos_x determine position of bar, p sets the number of asterisks, y_dist sets y distance of the asterisk to
def independence_test(df_test,
column_value="speed",
column_factor="city_id",
alpha=0.05,
decimals=6):
"""
Function that carries out indepence test given a numerical column and a factor column.
First, it is checked that assumptions for ANOVA test are fullfilled.
Second, if so, ANOVA test is carried out. If not, a non parametric test, Kruskal wallis is performed
Args: A dataframe, a numerical column, a factor column, an alpha factor for the alpha level of the test,
decimals to round the p-value
Returns: A dataframe with the results of the test
"""
##We are extracting all posible combinations:
n_cases = list(df_test[column_factor].unique())
tuple_combinations = [(x, y) for x in n_cases for y in n_cases
if n_cases.index(y) > n_cases.index(x)]
##Shapiro test to check the normal distribution of residuals
##Null hypothesis: data is drawn from normal distribution.
w_saphiro, pvalue_saphiro = stats.shapiro(df_test[column_value].values)
df_out = pd.DataFrame(tuple_combinations)
### Kurskal test is performed in any case, and use where anova is not usable.
Kruskal_test = [
kruskal(
df_test[df_test[column_factor] == combination[0]]
[column_value].values, df_test[
df_test[column_factor] == combination[1]][column_value].values)
for combination in tuple_combinations
]
Kruskal_pvalue = [round(x.pvalue, decimals) for x in Kruskal_test]
Kruskal_H = [x.statistic for x in Kruskal_test]
## if saphiro test is passed, we continue calculating Anova:
if pvalue_saphiro > 0.05:
print("Saphiro test p-value: %s" % round(pvalue_saphiro, decimals))
print(
"Null hypothesis cannot be rejected. We asume residuals are normally distributed"
)
##Now let's check every possible combinatio to perform Bartlet test:
Bartlett_test = [
stats.bartlett(
df_test[df_test[column_factor] == combination[0]]
[column_value], df_test[df_test[column_factor] ==
combination[1]][column_value])
for combination in tuple_combinations
]
##Extracting p-values and statistic values:
Bartlett_pvalue = [x.pvalue for x in Bartlett_test]
Bartlett_w = [x.statistic for x in Bartlett_test]
Anova_test = [
stats.f_oneway(
df_test[df_test[column_factor] ==
combination[0]][column_value].values,
df_test[df_test[column_factor] ==
combination[1]][column_value].values)
for combination in tuple_combinations
]
Anova_pvalue = [x.pvalue for x in Anova_test]
Anova_f = [x.statistic for x in Anova_test]
df_out = pd.concat([
df_out,
pd.DataFrame(Bartlett_pvalue),
pd.DataFrame(Anova_pvalue),
pd.DataFrame(Kruskal_pvalue)
],
axis=1)
df_out.columns = [
"First_value", "Second_value", "Bartlett_pvalue", "Anova_pvalue",
"kruskal_pvalue"
]
df_out["Reject Bartlett null hyp"] = np.where(
df_out["Bartlett_pvalue"] < alpha, True, False)
df_out["Reject Anova null hyp"] = np.where(
df_out["Anova_pvalue"] < alpha, True, False)
## Kruscal_pvalue
df_out["kruskal_pvalue"] = np.where(
df_out["Reject Bartlett null hyp"] == True,
df_out["kruskal_pvalue"], np.NaN)
else:
print("Saphiro test p-value: %s" % round(pvalue_saphiro, 4))
print(
"Null hypothesis rejected. We cannot asume residuals are normally distributed"
)
print("Using kruskal wallis test")
df_out = pd.concat(
[df_out,
pd.DataFrame(Kruskal_pvalue),
pd.DataFrame(Kruskal_H)],
axis=1)
# df_out=pd.concat([df_out,pd.DataFrame([Kruskal_H])],axis=1)
df_out.columns = [
"First_value", "Second_value", "kruskal_p_value", "kruskal_H_value"
]
df_out["Reject Kruskal null hyp"] = np.where(
df_out["kruskal_p_value"] < alpha, True, False)
return df_out