def estat(x, y, nboot=1000, replace=False, method='log', fitting=False):
'''
Energy distance statistics test.
Reference
---------
Aslan, B, Zech, G (2005) Statistical energy as a tool for binning-free
multivariate goodness-of-fit tests, two-sample comparison and unfolding.
Nuc Instr and Meth in Phys Res A 537: 626-636
Szekely, G, Rizzo, M (2014) Energy statistics: A class of statistics
based on distances. J Stat Planning & Infer 143: 1249-1272
Brian Lau, multdist, https://github.com/brian-lau/multdist
'''
n, N = len(x), len(x) + len(y)
stack = np.vstack([x, y])
stack = (stack - stack.mean(0)) / stack.std(0)
if replace:
rand = lambda x: random.randint(x, size=x)
else:
rand = random.permutation
en = energy(stack[:n], stack[n:], method)
en_boot = np.zeros(nboot, 'f')
for i in range(nboot):
idx = rand(N)
en_boot[i] = energy(stack[idx[:n]], stack[idx[n:]], method)
if fitting:
param = genextreme.fit(en_boot)
p = genextreme.sf(en, *param)
return p, en, param
else:
p = (en_boot >= en).sum() / nboot
return p, en, en_boot
def _compare_resamples(self, tvalues, null_max_tvalues, null_min_tvalues):
pvalues = []
maxparams = genextreme.fit(null_max_tvalues)
minparams = genextreme.fit([-x for x in null_min_tvalues])
for tvalue in tvalues:
pvalue = genextreme.sf(tvalue, *maxparams) if tvalue >= 0 else genextreme.sf(-tvalue, *minparams)
pvalues.append(pvalue)
return pvalues
def estat(x, y, nboot=1000, replace=False, method='log', fitting=False):
'''
Energy distance statistics test.
Reference
---------
Aslan, B, Zech, G (2005) Statistical energy as a tool for binning-free
multivariate goodness-of-fit tests, two-sample comparison and unfolding.
Nuc Instr and Meth in Phys Res A 537: 626-636
Szekely, G, Rizzo, M (2014) Energy statistics: A class of statistics
based on distances. J Stat Planning & Infer 143: 1249-1272
Brian Lau, multdist, https://github.com/brian-lau/multdist
'''
n, N = len(x), len(x) + len(y)
stack = np.vstack([x, y])
stack = (stack - stack.mean(0)) / stack.std(0)
if replace:
rand = lambda x: random.randint(x, size=x)
else:
rand = random.permutation
en = energy(stack[:n], stack[n:], method)
en_boot = np.zeros(nboot, 'f')
for i in range(nboot):
idx = rand(N)
en_boot[i] = energy(stack[idx[:n]], stack[idx[n:]], method)
if fitting:
param = genextreme.fit(en_boot)
p = genextreme.sf(en, *param)
return p, en, param
else:
p = (en_boot >= en).sum() / nboot
return p, en, en_boot
def extreme_values(weighted_residuals, confidence_interval):
'''
This function uses extreme value theory to calculate the number of
standard deviations away from the mean at which we should expect to bracket
*all* of our n data points at a certain confidence level.
It then uses that value to identify which (if any) of the data points
lie outside that region, and calculates the corresponding probabilities
of finding a data point at least that many standard deviations away.
Parameters
----------
weighted_residuals : array of floats
Array of residuals weighted by the square root of their
variances wr_i = r_i/sqrt(var_i)
confidence_interval : float
Probability at which all the weighted residuals lie
within the confidence bounds
Returns
-------
confidence_bound : float
Number of standard deviations at which we should expect to encompass
all data at the user-defined confidence interval.
indices : array of floats
Indices of weighted residuals exceeding the confidence_interval
defined by the user
probabilities : array of floats
The probabilities that the extreme data point of the distribution lies
further from the mean than the observed position wr_i for each i in
the "indices" output array.
'''
n = len(weighted_residuals)
mean = norm.isf(1./n)
# good approximation for > 10 data points
scale = 0.8/np.power(np.log(n), 1./2.)
# good approximation for > 10 data points
c = 0.33/np.power(np.log(n), 3./4.)
# We now need a 1-tailed probability from the given confidence_interval
# p_total = 1. - confidence_interval = p_upper + p_lower - p_upper*p_lower
# p_total = 1. - confidence_interval = 2p - p^2, therefore:
p = 1. - np.sqrt(confidence_interval)
confidence_bound = genextreme.isf(p, c, loc=mean, scale=scale)
indices = [i for i, r in enumerate(weighted_residuals)
if np.abs(r) > confidence_bound]
# Convert back to 2-tailed probabilities
probabilities = (1.
- np.power(genextreme.sf(np.abs(weighted_residuals[indices]),
c, loc=mean, scale=scale) - 1., 2.))
return confidence_bound, indices, probabilities
def estat(x,
y,
nboot=1000,
maxt=60.,
replace=False,
method='log',
fitting=False):
"""
Energy distance statistics test.
References
----------
* Aslan, B, Zech, G (2005) Statistical energy as a tool for binning-free
multivariate goodness-of-fit tests, two-sample comparison and unfolding.
Nuc Instr and Meth in Phys Res A 537: 626-636
* Szekely, G, Rizzo, M (2014) Energy statistics: A class of statistics
based on distances. J Stat Planning & Infer 143: 1249-1272
* Brian Lau, multdist, https://github.com/brian-lau/multdist
"""
n, N = len(x), len(x) + len(y)
stack = np.vstack([x, y])
# stack = (stack - stack.mean(0)) / stack.std(0)
stack = (stack - np.nanmean(stack, 0)) / np.nanstd(stack, 0)
if replace:
def rand(x):
return np.random.randint(x, size=x)
# rand = lambda x: np.random.randint(x, size=x)
else:
rand = np.random.permutation
en = energy(stack[:n], stack[n:], method)
en_boot = np.zeros(nboot, 'f')
s = t.time()
for i in range(nboot):
idx = rand(N)
en_boot[i] = energy(stack[idx[:n]], stack[idx[n:]], method)
if t.time() - s > maxt:
print("Time consumed, exit bootstrap (N={})".format(i))
en_boot, nboot = en_boot[:i], i + 1
break
if fitting:
param = genextreme.fit(en_boot)
p = genextreme.sf(en, *param)
return p, en, param
else:
p = (en_boot >= en).sum() / nboot
return p, en, en_boot
def extreme_values(weighted_residuals, confidence_interval):
'''
This function uses extreme value theory to calculate the number of
standard deviations away from the mean at which we should expect to bracket
*all* of our n data points at a certain confidence level.
It then uses that value to identify which (if any) of the data points
lie outside that region, and calculates the corresponding probabilities
of finding a data point at least that many standard deviations away.
Parameters
----------
weighted_residuals : array of floats
Array of residuals weighted by the square root of their
variances wr_i = r_i/sqrt(var_i)
confidence_interval : float
Probability at which all the weighted residuals lie
within the confidence bounds
Returns
-------
confidence_bound : float
Number of standard deviations at which we should expect to encompass all
data at the user-defined confidence interval.
indices : array of floats
Indices of weighted residuals exceeding the confidence_interval
defined by the user
probabilities : array of floats
The probabilities that the extreme data point of the distribution lies
further from the mean than the observed position wr_i for each i in
the "indices" output array.
'''
n=len(weighted_residuals)
mean = norm.isf(1./n)
scale = 0.8/np.power(np.log(n), 1./2.) # good approximation for > 10 data points
c = 0.33/np.power(np.log(n), 3./4.) # good approximation for > 10 data points
# We now need a 1-tailed probability from the given confidence_interval
# p_total = 1. - confidence_interval = p_upper + p_lower - p_upper*p_lower
# p_total = 1. - confidence_interval = 2p - p^2, therefore:
p = 1. - np.sqrt(confidence_interval)
confidence_bound = genextreme.isf(p, c, loc=mean, scale=scale)
indices = [i for i, r in enumerate(weighted_residuals) if np.abs(r) > confidence_bound]
probabilities = 1. - np.power(genextreme.sf(np.abs(weighted_residuals[indices]), c, loc=mean, scale=scale) - 1., 2.) # Convert back to 2-tailed probabilities
return confidence_bound, indices, probabilities
def calculate_adjusted_p_values_genextreme(in_master_table, c, loc, scale,
in_alpha):
raw_p_values = list()
for i in range(0, len(in_master_table)):
tmp_p_val = genextreme.sf(in_master_table[i][7], c, loc, scale)
raw_p_values.append(tmp_p_val)
master_table[i].append(tmp_p_val)
# adjust p-values
if len(raw_p_values) >= 2:
adjusted_p_values = multipletests(raw_p_values,
alpha=in_alpha,
method='fdr_bh',
is_sorted=False)
for i in range(0, len(adjusted_p_values[1])):
in_master_table[i].append(adjusted_p_values[1][i])
else:
for i in range(0, len(in_master_table)):
in_master_table[i].append("na")
return in_master_table
def estat(x, y, nboot=1000, replace=False, method='log', fitting=False):
"""
Energy distance statistics test.
Compares d-dimensional data from two samples using a measure based on
statistical energy. The test is non-parametric, does not require binning
and easily scales to arbitrary dimensions.
The analytic distribution of the statistic is unknown, and p-values
are estimated using a permutation procedure, which works well
according to simulations by Aslan & Zech.
INPUTS
x - [n1 x d] matrix
y - [n2 x d] matrix
OPTIONAL
flag - 'sr', Szekely & Rizzo energy statistic
'az', Aslan & Zech energy statistic (default)
nboot - # of bootstrap resamples (default = 1000)
replace - boolean for sampling with replacement (default = false)
OUTPUTS
p - p-value by permutation
e_n - minimum energy statistic
e_n_boot - bootstrap samples
References
----------
* Aslan, B, Zech, G (2005) Statistical energy as a tool for binning-free
multivariate goodness-of-fit tests, two-sample comparison and unfolding.
Nuc Instr and Meth in Phys Res A 537: 626-636
* Szekely, G, Rizzo, M (2014) Energy statistics: A class of statistics
based on distances. J Stat Planning & Infer 143: 1249-1272
* Brian Lau, multdist, https://github.com/brian-lau/multdist
"""
n, N = len(x), len(x) + len(y)
stack = np.vstack([x, y])
stack = (stack - stack.mean(0)) / stack.std(0)
if replace:
def rand(x):
return np.random.randint(x, size=x)
# rand = lambda x: np.random.randint(x, size=x)
else:
rand = np.random.permutation
en = energy(stack[:n], stack[n:], method)
en_boot = np.zeros(nboot, 'f')
for i in range(nboot):
idx = rand(N)
en_boot[i] = energy(stack[idx[:n]], stack[idx[n:]], method)
if fitting:
param = genextreme.fit(en_boot)
p = genextreme.sf(en, *param)
return p, en, param
else:
p = (en_boot >= en).sum() / nboot
return p, en, en_boot
