def fit(self, norm="zscore"):
""" Calculate the regression parameters and statistics. """
bold = self.bold.copy()
dm = self.dm.copy()
# Normalize both the bold and dm
if norm != None:
bold = self._normalize_array(bold, norm)
dm = self._normalize_array(dm, norm)
# Add movement regressors... if present
try:
dm_movement = self.data["movement"]
dm = np.vstack((dm, dm_movement))
except KeyError:
pass
# Append a dummy predictor and run the regression
#
# Dummy is added at the last minute so it does not
# interact with normalization or smoothing routines.
dm_dummy = np.ones((dm.shape[0], dm.shape[1] + 1))
dm_dummy[0:dm.shape[0], 0:dm.shape[1]] = dm
# Truncate bold or dm_dummy if needed, and Go!
try:
bold = bold[0:dm_dummy.shape[0]]
dm_dummy = dm_dummy[0:len(bold), :]
except IndexError:
pass
self.glm = GLS(bold, dm_dummy).fit()
def _orth_dm(self):
""" Orthgonalize (by regression) each col in self.dm with respect to
its left neighbor. """
dm = self.dm
## Rename for brevity
# Make sure conds and ncol dm
# are divisors
conds = list(set(self.trials))
nconds = len(conds) - 1 ## Drop baseline
ncols = dm.shape[1] - 1
if ncols % nconds:
raise ValueError(
"The number of condtions and shape of the dm are incompatible."
)
# If these are the same size there is nothing to
# orthgonalize.
if ncols != nconds:
orth_dm = np.zeros_like(dm)
orth_dm[:, 0] = dm[:, 0]
## Move baseline data over
# Use num_col_per_cond, along with nconds
# to find the strides we need to take along
# the DM to orthgonalize each set of col(s)
# belonging to each cond.
num_col_per_cond = ncols / nconds
for cond in conds:
# Skip baseline
if cond == 0:
continue
left = cond
right = cond + nconds
# Rolling loop over the cols_per_cond
# orthgonalizing as we go.
# Note: we never use cnt directly...
for cnt in range(num_col_per_cond - 1):
# Orthgonalize left col to right....
glm = GLS(dm[:, right], dm[:, left]).fit() ## GLS(y, x)
orth_dm[:, right] = glm.resid
orth_dm[:, left] = dm[:, left]
# Shift indices for next iteration.
left = deepcopy(right)
right = right + nconds
self.dm = orth_dm
else:
print("Nothing to orthgonalize.")
def gls_model(X, y, test_size=0.25):
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=test_size)
gls = GLS(y_train, X_train).fit(cov_type='HC1')
coeffs = (gls.params.to_frame().rename(columns={
0: "coeff"
}).merge(
gls.pvalues.to_frame().rename(columns={0: "p-value"}),
left_index=True,
right_index=True,
).assign(coeff=lambda df: df.apply(
lambda r: r["coeff"] if r["p-value"] < 0.05 else np.nan, axis=1))[[
"coeff"
]])
mse = gls.predict(X_test).pipe(lambda s: (s - y_test)**2).mean()
return {"model": gls, "coeffs": coeffs, "mse": mse}
