def test_predict_log_hazard_relative_to_mean_without_normalization(self, rossi):
cox = CoxPHFitter(normalize=False)
cox.fit(rossi, 'week', 'arrest')
log_relative_hazards = cox.predict_log_hazard_relative_to_mean(rossi)
means = rossi.mean(0).to_frame().T
assert cox.predict_partial_hazard(means).values[0][0] != 1.0
assert_frame_equal(log_relative_hazards, np.log(cox.predict_partial_hazard(rossi) / cox.predict_partial_hazard(means).squeeze()))
def test_using_dataframes_vs_numpy_arrays(self, data_pred2):
cf = CoxPHFitter()
cf.fit(data_pred2, 't', 'E')
X = data_pred2[cf.data.columns]
assert_frame_equal(
cf.predict_partial_hazard(np.array(X)),
cf.predict_partial_hazard(X)
)
def test_predict_log_hazard_relative_to_mean_with_normalization(self, rossi):
cox = CoxPHFitter(normalize=True)
cox.fit(rossi, 'week', 'arrest')
# they are equal because the data is normalized, so the mean of the covarites is all 0,
# thus exp(beta * 0) == 1, so exp(beta * X)/exp(beta * 0) = exp(beta * X)
assert_frame_equal(cox.predict_log_hazard_relative_to_mean(rossi), np.log(cox.predict_partial_hazard(rossi)))
def test_data_normalization(self, data_pred2):
# During fit, CoxPH copies the training data and normalizes it.
# Future calls should be normalized in the same way and
# internal training set should not be saved in a normalized state.
cf = CoxPHFitter()
cf.fit(data_pred2, duration_col='t', event_col='E')
# Internal training set
ci_trn = concordance_index(cf.durations,
-cf.predict_partial_hazard(cf.data).values,
cf.event_observed)
# New data should normalize in the exact same way
ci_org = concordance_index(data_pred2['t'],
-cf.predict_partial_hazard(data_pred2[['x1', 'x2']]).values,
data_pred2['E'])
assert ci_org == ci_trn
def test_data_normalization(self, data_pred2):
# During fit, CoxPH copies the training data and normalizes it.
# Future calls should be normalized in the same way and
# internal training set should not be saved in a normalized state.
cf = CoxPHFitter(normalize=True)
cf.fit(data_pred2, duration_col='t', event_col='E')
# Internal training set
ci_trn = concordance_index(cf.durations,
-cf.predict_partial_hazard(cf.data).values,
cf.event_observed)
# New data should normalize in the exact same way
ci_org = concordance_index(data_pred2['t'],
-cf.predict_partial_hazard(data_pred2[['x1', 'x2']]).values,
data_pred2['E'])
assert ci_org == ci_trn
def test_using_dataframes_vs_numpy_arrays(self, data_pred2):
# First without normalization
cf = CoxPHFitter(normalize=False)
cf.fit(data_pred2, 't', 'E')
X = data_pred2[cf.data.columns]
hazards = cf.predict_partial_hazard(X)
# A Numpy array should return the same result
hazards_n = cf.predict_partial_hazard(np.array(X))
assert np.all(hazards == hazards_n)
# Now with normalization
cf = CoxPHFitter(normalize=True)
cf.fit(data_pred2, 't', 'E')
hazards = cf.predict_partial_hazard(X)
# Compare with array argument
hazards_n = cf.predict_partial_hazard(np.array(X))
assert np.all(hazards == hazards_n)
def test_using_dataframes_vs_numpy_arrays(self, data_pred2):
# First without normalization
cf = CoxPHFitter(normalize=False)
cf.fit(data_pred2, 't', 'E')
X = data_pred2[cf.data.columns]
hazards = cf.predict_partial_hazard(X)
# A Numpy array should return the same result
hazards_n = cf.predict_partial_hazard(np.array(X))
assert np.all(hazards == hazards_n)
# Now with normalization
cf = CoxPHFitter(normalize=True)
cf.fit(data_pred2, 't', 'E')
hazards = cf.predict_partial_hazard(X)
# Compare with array argument
hazards_n = cf.predict_partial_hazard(np.array(X))
assert np.all(hazards == hazards_n)
def test_prediction_methods_respect_index(self, data_pred2):
x = data_pred2[['x1', 'x2']].ix[:3].sort_index(ascending=False)
expected_index = pd.Index(np.array([3, 2, 1, 0]))
cph = CoxPHFitter()
cph.fit(data_pred2, duration_col='t', event_col='E')
npt.assert_array_equal(cph.predict_partial_hazard(x).index, expected_index)
npt.assert_array_equal(cph.predict_percentile(x).index, expected_index)
npt.assert_array_equal(cph.predict_expectation(x).index, expected_index)
aaf = AalenAdditiveFitter()
aaf.fit(data_pred2, duration_col='t', event_col='E')
npt.assert_array_equal(aaf.predict_percentile(x).index, expected_index)
npt.assert_array_equal(aaf.predict_expectation(x).index, expected_index)
def test_cox_ph_prediction_monotonicity(self, data_pred2):
# Concordance wise, all prediction methods should be monotonic versions
# of one-another, unless numerical factors screw it up.
t = data_pred2['t']
e = data_pred2['E']
X = data_pred2[['x1', 'x2']]
cf = CoxPHFitter()
cf.fit(data_pred2, duration_col='t', event_col='E')
# Base comparison is partial_hazards
ci_ph = concordance_index(t, -cf.predict_partial_hazard(X).values, e)
ci_med = concordance_index(t, cf.predict_median(X).ravel(), e)
assert ci_ph == ci_med
ci_exp = concordance_index(t, cf.predict_expectation(X).ravel(), e)
assert ci_ph == ci_exp
def test_cox_ph_prediction_monotonicity(self, data_pred2):
# Concordance wise, all prediction methods should be monotonic versions
# of one-another, unless numerical factors screw it up.
t = data_pred2['t']
e = data_pred2['E']
X = data_pred2[['x1', 'x2']]
for normalize in [True, False]:
msg = ("Predict methods should get the same concordance" +
" when {}normalizing".format('' if normalize else 'not '))
cf = CoxPHFitter(normalize=normalize)
cf.fit(data_pred2, duration_col='t', event_col='E')
# Base comparison is partial_hazards
ci_ph = concordance_index(t, -cf.predict_partial_hazard(X).values, e)
ci_med = concordance_index(t, cf.predict_median(X).ravel(), e)
assert ci_ph == ci_med, msg
ci_exp = concordance_index(t, cf.predict_expectation(X).ravel(), e)
assert ci_ph == ci_exp, msg
def test_cox_ph_prediction_monotonicity(self, data_pred2):
# Concordance wise, all prediction methods should be monotonic versions
# of one-another, unless numerical factors screw it up.
t = data_pred2['t']
e = data_pred2['E']
X = data_pred2[['x1', 'x2']]
for normalize in [True, False]:
msg = ("Predict methods should get the same concordance" +
" when {}normalizing".format('' if normalize else 'not '))
cf = CoxPHFitter(normalize=normalize)
cf.fit(data_pred2, duration_col='t', event_col='E')
# Base comparison is partial_hazards
ci_ph = concordance_index(t, -cf.predict_partial_hazard(X).values, e)
ci_med = concordance_index(t, cf.predict_median(X).ravel(), e)
assert ci_ph == ci_med, msg
ci_exp = concordance_index(t, cf.predict_expectation(X).ravel(), e)
assert ci_ph == ci_exp, msg
def test_concordance_index_fast_is_same_as_slow():
size = 100
T = np.random.normal(size=size)
P = np.random.normal(size=size)
C = np.random.choice([0, 1], size=size)
Z = np.zeros_like(T)
# Hard to imagine these failing
assert slow_cindex(T, Z, C) == fast_cindex(T, Z, C)
assert slow_cindex(T, T, C) == fast_cindex(T, T, C)
# This is the real test though
assert slow_cindex(T, P, C) == fast_cindex(T, P, C)
cp = CoxPHFitter()
df = load_rossi()
cp.fit(df, duration_col='week', event_col='arrest')
T = cp.durations.values.ravel()
P = -cp.predict_partial_hazard(cp.data).values.ravel()
E = cp.event_observed.values.ravel()
assert slow_cindex(T, P, E) == fast_cindex(T, P, E)
def test_concordance_index_fast_is_same_as_slow():
size = 100
T = np.random.normal(size=size)
P = np.random.normal(size=size)
C = np.random.choice([0, 1], size=size)
Z = np.zeros_like(T)
# Hard to imagine these failing
assert slow_cindex(T, Z, C) == fast_cindex(T, Z, C)
assert slow_cindex(T, T, C) == fast_cindex(T, T, C)
# This is the real test though
assert slow_cindex(T, P, C) == fast_cindex(T, P, C)
cp = CoxPHFitter()
df = load_rossi()
cp.fit(df, duration_col='week', event_col='arrest')
T = cp.durations.values.ravel()
P = -cp.predict_partial_hazard(cp.data).values.ravel()
E = cp.event_observed.values.ravel()
assert slow_cindex(T, P, E) == fast_cindex(T, P, E)
def test_predict_log_hazard_relative_to_mean(self, rossi):
cox = CoxPHFitter()
cox.fit(rossi, 'week', 'arrest')
log_relative_hazards = cox.predict_log_hazard_relative_to_mean(rossi)
means = rossi.mean(0).to_frame().T
assert_frame_equal(log_relative_hazards, np.log(cox.predict_partial_hazard(rossi) / cox.predict_partial_hazard(means).squeeze()))
