def test_PE_F(self):
for adaptive_weights in [False, 'ridge', 'decisiontree', 'ridgeCV','decisiontreeCV']: #
is_correct = 1
try:
model = None
if adaptive_weights == False:
df, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
holdout, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
covar_importance = np.array([4,3,2,1,0,0,0])
weight_array = covar_importance/covar_importance.sum()
model = matching.FLAME(repeats=False, verbose=0,adaptive_weights =adaptive_weights)
model.fit(holdout_data=holdout,weight_array = list(weight_array))
output = model.predict(df)
else:
df, true_TE = generate_uniform_given_importance()
holdout, true_TE = generate_uniform_given_importance()
model = matching.FLAME(repeats=False, verbose=0,adaptive_weights =adaptive_weights)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct,
msg='FLAME-Error when we use PE method: {0} '.format(str(adaptive_weights)))
def test_verbose_F(self):
#Test verbose
df, true_TE = generate_uniform_given_importance()
for verbose in [0,1,2,3]:
is_correct = 1
try:
df, true_TE = generate_uniform_given_importance(num_control=1000, num_treated=1000,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
holdout, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
covar_importance = np.array([4,3,2,1,0,0,0])
weight_array = covar_importance/covar_importance.sum()
model = matching.FLAME(missing_data_replace = 2, want_bf = True, verbose = verbose)
model.fit(holdout_data=holdout)
output = model.predict(df)
model = matching.FLAME(repeats=True,verbose=verbose)
model.fit(holdout_data=0.5)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
break
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct, msg='FLAME-Error when verbose ={0}'.format(verbose))
def test_pre_dame_F(self):
df, true_TE = generate_uniform_given_importance(num_control=500, num_treated=500,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
holdout, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
is_correct = 1
try:
covar_importance = np.array([4,3,2,1,0,0,0])
weight_array = covar_importance/covar_importance.sum()
for x in [False, True]:
for y in [False, True]:
model1 = matching.FLAME(repeats=x,want_pe = y, want_bf = y,verbose=0,adaptive_weights = False)
model1.fit(holdout_data=holdout,weight_array = list(weight_array))
output = model1.predict(df, pre_dame = True)
model2 = matching.FLAME(repeats=x, want_pe = y, want_bf = y,verbose=0,adaptive_weights = 'decisiontreeCV')
model2.fit(holdout_data=holdout)
output = model2.predict(df, pre_dame = True)
output = model2.predict(df, pre_dame = True)
if check_statistics(model1) or check_statistics(model2) :
is_correct = 0
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct,
msg='FLAME-Error when we use pre_dame')
def test_miss_data_indicator_F(self):
is_correct = 1
try:
df, true_TE = generate_uniform_given_importance(num_control=1000,
num_treated=1000)
#Create missing df
m, n = df.shape
for i in range(int(m / 100)):
for j in [0, int(n / 2)]:
df.iloc[i, j] = 'a'
holdout = df.copy()
model = matching.FLAME(missing_indicator='a',
missing_holdout_replace=1,
missing_data_replace=1)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1,
is_correct,
msg='FLAME-Error when missing_indicator')
def broken_treatment_column_name_value():
df, true_TE = generate_uniform_given_importance()
holdout, true_TE = generate_uniform_given_importance()
df.loc[0,'treated'] = 4
model = matching.FLAME()
model.fit(holdout_data=holdout)
output = model.predict(df)
def test_large_C_repeats_F(self):
df_path = os.path.join((os.path.dirname(__file__)), 'basicTestData.csv')
df = pd.read_csv(df_path)
holdout_path = os.path.join((os.path.dirname(__file__)), 'basicHoldoutData.csv')
holdout = pd.read_csv(holdout_path)
model = matching.FLAME(repeats=False, verbose=1)
model.fit(holdout_data=holdout)
algo_output = model.predict(df, C=100000)
result_path = os.path.join((os.path.dirname(__file__)), 'basicResultData.csv')
result = pd.read_csv(result_path, index_col="Unnamed: 0")
dfs_equal = 1
try:
for index in result.index:
for col in result.columns:
if (result.loc[index, col] != "*" and
algo_output.loc[index, col] != "*" and
int(result.loc[index, col]) != int(algo_output.loc[index, col])):
print("index, col", index, col)
dfs_equal = 0
break
except (KeyError, ValueError):
# We would hit this block if theres a key error, so df columns
# are not equal or have different units, or weird entry in df, (string)
dfs_equal = 0
self.assertEqual(1, dfs_equal,
msg='Data frames not equal on index {0}, col {1}'.format(index, col))
def broken_holdout_type():
df, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100)
holdout = df.copy()
holdout.iloc[0,0] = 's'
model = matching.FLAME()
model.fit(holdout_data=holdout)
output = model.predict(df)
def test_miss_data_F(self):
is_correct = 1
try:
for missing_holdout_replace in [0, 1, 2]:
for missing_data_replace in [0, 1, 2, 3]:
df, true_TE = generate_uniform_given_importance(
num_control=1000, num_treated=1000)
#Create missing df
m, n = df.shape
for i in range(int(m / 100)):
for j in [0, int(n / 2)]:
df.iloc[i, j] = np.nan
holdout = df.copy()
model = matching.FLAME(
missing_holdout_replace=missing_holdout_replace,
missing_data_replace=missing_data_replace)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
break
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct, msg='FLAME-Error when do missing data'\
'handling with missing_holdout_replace = {0},missing_data_replace{1}'.format(missing_holdout_replace,missing_data_replace))
def test_datasets_F(self):
df_path = os.path.join((os.path.dirname(__file__)), 'basicTestData.csv')
for gen in [generate_uniform_given_importance,generate_binomial_given_importance,generate_binomial_decay_importance,df_path]:
is_correct = 1
try:
df = None
holdout = None
if type(gen) != str:
df, true_TE = gen()
holdout, true_TE = gen()
else:
df = gen
holdout = gen
model = matching.FLAME(repeats=False)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
break
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct,
msg='FLAME-Error when we use the dataset generated by {0} '.format(str(gen)))
def broken_C():
df, true_TE = generate_uniform_given_importance()
holdout, true_TE = generate_uniform_given_importance()
model = matching.FLAME()
model.fit(holdout_data=holdout)
output = model.predict(df,C = -1)
def test_repeats_F(self):
#Test other parameters
df, true_TE = generate_uniform_given_importance(
num_control=100,
num_treated=100,
num_cov=7,
min_val=0,
max_val=3,
covar_importance=[4, 3, 2, 1, 0, 0, 0])
holdout, true_TE = generate_uniform_given_importance(
num_control=100,
num_treated=100,
num_cov=7,
min_val=0,
max_val=3,
covar_importance=[4, 3, 2, 1, 0, 0, 0])
is_correct = 1
try:
model = matching.FLAME(repeats=True)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct, msg='FLAME-Error when repeat = True')
def broken_false_dataset():
df, true_TE = generate_uniform_given_importance(num_control=1000,
num_treated=1000)
holdout, true_TE = generate_uniform_given_importance(
num_control=1000, num_treated=1000)
model = matching.FLAME()
model.fit(holdout_data=holdout)
output = model.predict(False)
def broken_data_len():
df, true_TE = generate_uniform_given_importance(num_control=1000, num_treated=1000,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
holdout, true_TE = generate_uniform_given_importance()
model = matching.FLAME()
model.fit(holdout_data=holdout)
output = model.predict(df)
def test_other_param_F(self):
is_correct = 1
try:
df, true_TE = generate_uniform_given_importance(num_control=1000, num_treated=1000,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
holdout, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
model = matching.FLAME( early_stop_pe= 1, verbose=0)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
# model = matching.FLAME( stop_unmatched_c= True, verbose=0)
# model.fit(holdout_data=holdout)
# output = model.predict(df)
# if check_statistics(model):
# is_correct = 0
# model = matching.FLAME(stop_unmatched_t= True, verbose=0)
# model.fit(holdout_data=holdout)
# output = model.predict(df)
# if check_statistics(model):
# is_correct = 0
model = matching.FLAME(early_stop_un_c_frac = 0.5, verbose=0)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
model = matching.FLAME(early_stop_un_t_frac = 0.5, verbose=0)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
model = matching.FLAME(early_stop_iterations= 2, verbose=0)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct, msg='FLAME-Error when other parameters')
def broken_column_match():
df, true_TE = generate_uniform_given_importance()
holdout, true_TE = generate_uniform_given_importance()
set_ = holdout.columns
set_ = list(set_)
set_[0] = 'dasfadf'
holdout.columns = set_
model = matching.FLAME()
model.fit(holdout_data=holdout)
output = model.predict(df)
def broken_missing_data_replace():
df, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
holdout, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
covar_importance = np.array([4,3,2,1,0,0,0])
weight_array = covar_importance/covar_importance.sum()
model = matching.FLAME(missing_data_replace = 2, adaptive_weights =False)
model.fit(holdout_data=holdout,weight_array = list(weight_array))
output = model.predict(df)
def test_no_matching_F(self):
#Test data split
df, true_TE = generate_uniform_given_importance(num_control=3000, num_treated=3000)
is_correct = 1
try:
df = pd.DataFrame([[1,2,0,1.0],[3,4,1,2.0],[5,6,0,5.0],[7,8,1,8.0],[9,10,1,10.0]])
df.columns = ['cov1','cov2','treated','outcome']
holdout = df.copy()
model = matching.FLAME(repeats=True)
model.fit(holdout_data=holdout)
output = model.predict(df)
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct, msg='FLAME-Error when no matching')
def test_data_split_F(self):
#Test data split
df, true_TE = generate_uniform_given_importance(num_control=3000, num_treated=3000)
is_correct = 1
try:
for holdout in [0.3,0.5,0.7]:
model = matching.FLAME(repeats=True)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
break
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct, msg='FLAME-Error when holdout = {0}'.format(holdout))
def test_has_unmatched_units_F(self):
#Test data split
is_correct = 1
try:
df = pd.DataFrame([[1,2,0,1.0],[3,4,0,2.0],[5,6,0,5.0],[7,8,0,8.0],[9,10,1,10.0],[9,20,0,10.0]])
df.columns = ['cov1','cov2','treated','outcome']
holdout = df.copy()
model = matching.FLAME(repeats=True)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model):
is_correct = 0
except (KeyError, ValueError):
is_correct = 0
self.assertEqual(1, is_correct, msg='FLAME-Error when no matching')
def test_want_pebf_F(self):
#Test
df, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
holdout, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100,
num_cov=7, min_val=0,
max_val=3, covar_importance=[4,3,2,1,0,0,0])
is_correct = 1
try:
for want_pe in [False, True]:
for want_bf in [False, True]:
model = matching.FLAME(want_pe=want_pe,want_bf=want_bf)
model.fit(holdout_data=holdout)
output = model.predict(df)
if check_statistics(model) or (want_pe and len(model.pe_each_iter)==0) or (want_bf and len(model.bf_each_iter)==0):
is_correct = 0
break
except (KeyError, ValueError):
is_wrong = 0
self.assertEqual(1, is_correct, msg='FLAME Error when want_pe = {0} want_bf = {1}'.format(str(want_pe),str(want_bf)))
def broken_early_stop_pe_frac():
df, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100)
model = matching.FLAME(early_stop_pe_frac = 123)
model.fit(holdout_data=df)
output = model.predict(df)
def broken_outcome_column_name():
df, true_TE = generate_uniform_given_importance()
holdout, true_TE = generate_uniform_given_importance()
model = matching.FLAME()
model.fit(holdout_data=holdout,outcome_column_name = "sadfdag")
output = model.predict(df)
def broken_ATE_input_model():
model = matching.FLAME()
ATE(model)
def broken_weight_array_sum():
df, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100)
model = matching.FLAME(adaptive_weights = False)
model.fit(holdout_data=df, weight_array = [1,1,1,1])
output = model.predict(df)
def broken_adaptive_weights():
df, true_TE = generate_uniform_given_importance(num_control=100, num_treated=100)
model = matching.FLAME(adaptive_weights = 'safdsaf')
model.fit(holdout_data=df)
output = model.predict(df)