def ihdp_miwae(set_id_range=range(1, 1001),
prop_miss=0.1,
seed=0,
d_miwae=3,
n_epochs=602,
sig_prior=1,
add_wy=False,
method="glm",
**kwargs):
from miwae import miwae
l_scores = []
for set_id in set_id_range:
X = pd.read_csv('./data/IHDP/csv/R_ate_ihdp_npci_' + str(set_id) +
'.csv')
w = np.array(X.iloc[:, 0]).reshape((-1, 1))
y = np.array(X.iloc[:, 1]).reshape((-1, 1))
X = np.array(X.iloc[:, 5:])
X_miss = ampute(X, prop_miss=prop_miss, seed=seed)
if set_id == 1:
if add_wy:
xhat, zhat, zhat_mul = miwae(X_miss,
d_miwae=d_miwae,
sig_prior=sig_prior,
n_epochs=n_epochs,
add_wy=add_wy,
w=w,
y=y)
else:
xhat, zhat, zhat_mul = miwae(X_miss,
d_miwae=d_miwae,
sig_prior=sig_prior,
n_epochs=n_epochs,
add_wy=add_wy)
# print('shape of outputs miwae:')
# print('xhat.shape, zhat.shape, zhat_mul.shape:')
# (1000, 200) (1000, 3) (200, 1000, 3)
print(xhat.shape, zhat.shape, zhat_mul.shape)
print(set_id)
# Tau estimated on Zhat=E[Z|X]
ps_hat, y0_hat, y1_hat = get_ps_y01_hat(zhat, w, y)
res_tau_ols = tau_ols(zhat, w, y)
res_tau_ols_ps = tau_ols_ps(zhat, w, y)
res_tau_dr = tau_dr(y, w, y0_hat, y1_hat, ps_hat, method)
res_tau_diffmu = np.mean(y1_hat - y0_hat)
lr = LinearRegression()
lr.fit(zhat, y)
y_hat = lr.predict(zhat)
res_tau_resid = tau_residuals(y, w, y_hat, ps_hat, method)
# Tau estimated on Zhat^(b), l=1,...,B sampled from posterior
res_mul_tau_dr = []
res_mul_tau_ols = []
res_mul_tau_ols_ps = []
res_mul_tau_resid = []
res_mul_tau_diffmu = []
for zhat_b in zhat_mul:
ps_hat, y0_hat, y1_hat = get_ps_y01_hat(zhat_b, w, y)
res_mul_tau_dr.append(tau_dr(y, w, y0_hat, y1_hat, ps_hat, method))
res_mul_tau_ols.append(tau_ols(zhat_b, w, y))
res_mul_tau_ols_ps.append(tau_ols_ps(zhat_b, w, y))
res_mul_tau_diffmu.append(np.mean(y1_hat - y0_hat))
lr = LinearRegression()
lr.fit(zhat, y)
y_hat = lr.predict(zhat_b)
res_mul_tau_resid = tau_residuals(y, w, y_hat, ps_hat, method)
res_mul_tau_dr = np.mean(res_mul_tau_dr)
res_mul_tau_ols = np.mean(res_mul_tau_ols)
res_mul_tau_ols_ps = np.mean(res_mul_tau_ols_ps)
res_mul_tau_resid = np.mean(res_mul_tau_resid)
res_mul_tau_diffmu = np.mean(res_mul_tau_diffmu)
dcor_zhat = np.nan
dcor_zhat_mul = np.nan
score = [
prop_miss, d_miwae, sig_prior, n_epochs, add_wy, set_id,
res_tau_dr, res_tau_ols, res_tau_ols_ps, res_tau_resid,
res_tau_diffmu, res_mul_tau_dr, res_mul_tau_ols,
res_mul_tau_ols_ps, res_mul_tau_resid, res_mul_tau_diffmu,
dcor_zhat, dcor_zhat_mul
]
l_scores.append(score)
score_data = pd.DataFrame(
l_scores,
columns=list([
'prop_miss', 'd_miwae', 'sig_prior', 'n_epochs', 'add_wy',
'set_id', 'res_tau_dr', 'res_tau_ols', 'res_tau_ols_ps',
'res_tau_resid', 'res_tau_diffmu', 'res_mul_tau_dr',
'res_mul_tau_ols', 'res_mul_tau_ols_ps', 'res_mul_tau_resid',
'res_mul_tau_diffmu', 'dcor_zhat', 'dcor_zhat_mul'
]))
return score_data
def exp_miwae(model="dlvm",
n=1000,
d=3,
p=100,
prop_miss=0.1,
citcio=False,
seed=0,
d_miwae=3,
n_epochs=602,
sig_prior=1,
add_wy=False,
num_samples_zmul=200,
method="glm",
**kwargs):
from miwae import miwae
if model == "lrmf":
Z, X, w, y, ps = gen_lrmf(n=n,
d=d,
p=p,
citcio=citcio,
prop_miss=prop_miss,
seed=seed)
elif model == "dlvm":
Z, X, w, y, ps = gen_dlvm(n=n,
d=d,
p=p,
citcio=citcio,
prop_miss=prop_miss,
seed=seed)
else:
raise NotImplementedError(
"Other data generating models not implemented here yet.")
X_miss = ampute(X, prop_miss=prop_miss, seed=seed)
if add_wy:
xhat, zhat, zhat_mul = miwae(X_miss,
d=d_miwae,
sig_prior=sig_prior,
num_samples_zmul=num_samples_zmul,
n_epochs=n_epochs,
add_wy=add_wy,
w=w,
y=y)
else:
xhat, zhat, zhat_mul = miwae(X_miss,
d=d_miwae,
sig_prior=sig_prior,
num_samples_zmul=num_samples_zmul,
n_epochs=n_epochs,
add_wy=add_wy)
# print('shape of outputs miwae:')
# print('xhat.shape, zhat.shape, zhat_mul.shape:')
# (1000, 200) (1000, 3) (200, 1000, 3)
print(xhat.shape, zhat.shape, zhat_mul.shape)
# Tau estimated on Zhat=E[Z|X]
ps_hat, y0_hat, y1_hat = get_ps_y01_hat(zhat, w, y)
res_tau_ols = tau_ols(zhat, w, y)
res_tau_ols_ps = tau_ols_ps(zhat, w, y)
res_tau_dr = tau_dr(y, w, y0_hat, y1_hat, ps_hat, method)
lr = LinearRegression()
lr.fit(zhat, y)
y_hat = lr.predict(zhat)
res_tau_resid = tau_residuals(y, w, y_hat, ps_hat, method)
# Tau estimated on Zhat^(b), l=1,...,B sampled from posterior
res_mul_tau_dr = []
res_mul_tau_ols = []
res_mul_tau_ols_ps = []
res_mul_tau_resid = []
for zhat_b in zhat_mul:
ps_hat, y0_hat, y1_hat = get_ps_y01_hat(zhat_b, w, y)
res_mul_tau_dr.append(tau_dr(y, w, y0_hat, y1_hat, ps_hat, method))
res_mul_tau_ols.append(tau_ols(zhat_b, w, y))
res_mul_tau_ols_ps.append(tau_ols_ps(zhat_b, w, y))
lr = LinearRegression()
lr.fit(zhat_b, y)
y_hat = lr.predict(zhat_b)
res_mul_tau_resid.append(tau_residuals(y, w, y_hat, ps_hat, method))
res_mul_tau_dr = np.mean(res_mul_tau_dr)
res_mul_tau_ols = np.mean(res_mul_tau_ols)
res_mul_tau_ols_ps = np.mean(res_mul_tau_ols_ps)
res_mul_tau_resid = np.mean(res_mul_tau_resid)
if Z.shape[1] == zhat.shape[1]:
dcor_zhat = dcor(Z, zhat)
dcor_zhat_mul = []
for zhat_b in zhat_mul:
dcor_zhat_mul.append(dcor(Z, zhat_b))
dcor_zhat_mul = np.mean(dcor_zhat_mul)
return res_tau_dr, res_tau_ols, res_tau_ols_ps, res_tau_resid, res_mul_tau_dr, res_mul_tau_ols, res_mul_tau_ols_ps, res_mul_tau_resid, dcor_zhat, dcor_zhat_mul
def ihdp_baseline(set_id=1,
prop_miss=0.1,
seed=0,
full_baseline=False,
add_wy=False,
sig_prior=1,
d_miwae=10,
n_epochs=10,
method="glm",
**kwargs):
X = pd.read_csv('./data/IHDP/csv/R_ate_ihdp_npci_' + str(set_id) + '.csv')
w = np.array(X.iloc[:, 0]).reshape((-1, 1))
y = np.array(X.iloc[:, 1]).reshape((-1, 1))
X = np.array(X.iloc[:, 5:])
X_miss = ampute(X, prop_miss=prop_miss, seed=seed)
X_imp_mean = np.zeros(X_miss.shape)
X_imp_mice = np.zeros(X_miss.shape)
try:
from sklearn.impute import SimpleImputer
X_imp_mean = SimpleImputer().fit_transform(X_miss)
except:
pass
try:
from sklearn.impute import IterativeImputer
X_imp_mice = IterativeImputer()().fit_transform(X_miss)
except:
pass
algo_name = ['X', 'X_imp_mean']
algo_ = [X, X_imp_mean]
if full_baseline:
# complete the baseline
Z_mf = get_U_softimpute(X_miss)
# need try-except for sklearn version
try:
from sklearn.impute import IterativeImputer
X_imp = IterativeImputer().fit_transform(X_miss)
except:
from sklearn.experimental import enable_iterative_imputer
from sklearn.impute import IterativeImputer
X_imp = IterativeImputer().fit_transform(X_miss)
algo_name += ['X_imp', 'Z_mf'] #, 'Z_perm']
algo_ += [X_imp, Z_mf] #, Z_perm]
tau = dict()
for name, zhat in zip(algo_name, algo_):
if name == 'X_mi':
res_tau_dr, res_tau_ols, res_tau_ols_ps, res_tau_resid = tau_mi(
zhat, w, y, method=method)
else:
ps_hat, y0_hat, y1_hat = get_ps_y01_hat(zhat, w, y)
res_tau_ols = tau_ols(zhat, w, y)
res_tau_ols_ps = tau_ols_ps(zhat, w, y)
res_tau_dr = tau_dr(y, w, y0_hat, y1_hat, ps_hat, method)
lr = LinearRegression()
lr.fit(zhat, y)
y_hat = lr.predict(zhat)
res_tau_resid = tau_residuals(y, w, y_hat, ps_hat, method)
tau[name] = res_tau_dr, res_tau_ols, res_tau_ols_ps, res_tau_resid
return tau
def exp_baseline(model="dlvm",
n=1000,
d=3,
p=100,
prop_miss=0.1,
citcio=False,
seed=0,
full_baseline=False,
method="glm",
**kwargs):
if model == "lrmf":
Z, X, w, y, ps = gen_lrmf(n=n,
d=d,
p=p,
citcio=citcio,
prop_miss=prop_miss,
seed=seed)
elif model == "dlvm":
Z, X, w, y, ps = gen_dlvm(n=n,
d=d,
p=p,
citcio=citcio,
prop_miss=prop_miss,
seed=seed)
else:
raise NotImplementedError(
"Other data generating models not implemented here yet.")
X_miss = ampute(X, prop_miss=prop_miss, seed=seed)
from sklearn.impute import SimpleImputer
X_imp_mean = SimpleImputer().fit_transform(X_miss)
Z_perm = np.random.permutation(Z)
# Z_rnd = np.random.randn(Z.shape[0], Z.shape[1])
algo_name = ['Z', 'X'] #, 'X_imp_mean']
algo_ = [Z, X] #, X_imp_mean]
if full_baseline:
# complete the baseline
Z_mf = get_U_softimpute(X_miss)
# need try-except for sklearn version
try:
from sklearn.impute import IterativeImputer
X_imp = IterativeImputer().fit_transform(X_miss)
except:
from sklearn.experimental import enable_iterative_imputer
from sklearn.impute import IterativeImputer
X_imp = IterativeImputer().fit_transform(X_miss)
algo_name += ['Z_mf'] #['X_imp','Z_mf']#, 'Z_perm']
algo_ += [Z_mf] #[X_imp, Z_mf]#, Z_perm]
tau = dict()
for name, zhat in zip(algo_name, algo_):
if name == 'X_mi':
res_tau_dr, res_tau_ols, res_tau_ols_ps, res_tau_resid = tau_mi(
zhat, w, y, method=method)
else:
ps_hat, y0_hat, y1_hat = get_ps_y01_hat(zhat, w, y)
res_tau_ols = tau_ols(zhat, w, y)
res_tau_ols_ps = tau_ols_ps(zhat, w, y)
res_tau_dr = tau_dr(y, w, y0_hat, y1_hat, ps_hat, method)
lr = LinearRegression()
lr.fit(zhat, y)
y_hat = lr.predict(zhat)
res_tau_resid = tau_residuals(y, w, y_hat, ps_hat, method)
tau[name] = res_tau_dr, res_tau_ols, res_tau_ols_ps, res_tau_resid
return tau