def __init__(self, y, w, permutations=PERMUTATIONS):
y = np.asarray(y).flatten()
self.n = len(y)
self.y = y
w.transform = "B"
self.w = w
self.permutations = permutations
self.__moments()
self.y2 = y * y
y = y.reshape(
len(y), 1
) # Ensure that y is an n by 1 vector, otherwise y*y.T == y*y
self.den_sum = (y * y.T).sum() - (y * y).sum()
self.G = self.__calc(self.y)
self.z_norm = (self.G - self.EG) / np.sqrt(self.VG)
self.p_norm = 1.0 - stats.norm.cdf(np.abs(self.z_norm))
if permutations:
sim = [
self.__calc(np.random.permutation(self.y)) for i in range(permutations)
]
self.sim = sim = np.array(sim)
above = sim >= self.G
larger = sum(above)
if (self.permutations - larger) < larger:
larger = self.permutations - larger
self.p_sim = (larger + 1.0) / (permutations + 1.0)
self.EG_sim = sum(sim) / permutations
self.seG_sim = sim.std()
self.VG_sim = self.seG_sim ** 2
self.z_sim = (self.G - self.EG_sim) / self.seG_sim
self.p_z_sim = 1.0 - stats.norm.cdf(np.abs(self.z_sim))
def calc(self):
w = self.w
W = w.sparse
self.y_sum = self.y.sum()
y = self.y
remove_self = not self.star
N = self.w.n - remove_self
statistic = (W @ y) / (y.sum() - y * remove_self)
# ----------------------------------------------------#
# compute moments necessary for analytical inference #
# ----------------------------------------------------#
empirical_mean = (y.sum() - y * remove_self) / N
# variance looks complex, yes, but it obtains from E[x^2] - E[x]^2.
# So, break it down to allow subtraction of the self-neighbor.
mean_of_squares = ((y**2).sum() - (y**2) * remove_self) / N
empirical_variance = mean_of_squares - empirical_mean**2
# Since we have corrected the diagonal, this should work
cardinality = np.asarray(W.sum(axis=1)).squeeze()
expected_value = cardinality / N
expected_variance = (cardinality * (N - cardinality) / (N - 1) *
(1 / N**2) * (empirical_variance /
(empirical_mean**2)))
z_scores = (statistic - expected_value) / np.sqrt(expected_variance)
self.Gs = statistic
self.EGs = expected_value
self.VGs = expected_variance
self.Zs = z_scores
def calc(self):
y = self.y
y2 = y * y
self.y_sum = y_sum = sum(y)
y2_sum = sum(y2)
if not self.star:
yl = 1.0 * slag(self.w, y)
ydi = y_sum - y
self.Gs = yl / ydi
N = self.n - 1
yl_mean = ydi / N
s2 = (y2_sum - y2) / N - (yl_mean) ** 2
else:
self.w.transform = "B"
yl = 1.0 * slag(self.w, y)
yl += y
if self.w_transform == "r":
yl = yl / (self.__getCardinalities() + 1.0)
self.Gs = yl / y_sum
N = self.n
yl_mean = y.mean()
s2 = y.var()
EGs_num, VGs_num = 1.0, 1.0
if self.w_transform == "b":
W = self.__getCardinalities()
W += self.star
EGs_num = W * 1.0
VGs_num = (W * (1.0 * N - W)) / (1.0 * N - 1)
self.EGs = (EGs_num * 1.0) / N
self.VGs = (VGs_num) * (1.0 / (N ** 2)) * ((s2 * 1.0) / (yl_mean ** 2))
self.Zs = (self.Gs - self.EGs) / np.sqrt(self.VGs)
self.w.transform = self.w_original
