def test_testers():
# Smoke test
np.random.seed(2432)
n = 200
p = 50
y = np.random.normal(size=n)
x = np.random.normal(size=(n, p))
testers = [
kr.CorrelationEffects(),
kr.ForwardEffects(pursuit=False),
kr.ForwardEffects(pursuit=True),
kr.OLSEffects()
]
for method in "equi", "sdp":
if method == "sdp" and not has_cvxopt:
continue
for tv in testers:
RegressionFDR(y, x, tv, design_method=method)
def test_sim():
# This function assesses the performance of the knockoff approach
# relative to its theoretical claims.
np.random.seed(43234)
npos = 30
target_fdr = 0.2
nrep = 10
testers = [[kr.CorrelationEffects(), 300, 100, 6],
[kr.ForwardEffects(pursuit=False), 300, 100, 3.5],
[kr.ForwardEffects(pursuit=True), 300, 100, 3.5],
[kr.OLSEffects(), 3000, 200, 3.5]]
for method in "equi", "sdp":
if method == "sdp" and not has_cvxopt:
continue
for tester_info in testers:
fdr = 0
power = 0
tester = tester_info[0]
n = tester_info[1]
p = tester_info[2]
es = tester_info[3]
for k in range(nrep):
x = np.random.normal(size=(n, p))
x /= np.sqrt(np.sum(x * x, 0))
coeff = es * (-1)**np.arange(npos)
y = np.dot(x[:, 0:npos], coeff) + np.random.normal(size=n)
kn = RegressionFDR(y, x, tester)
tr = kn.threshold(target_fdr)
cp = np.sum(kn.stats >= tr)
cp = max(cp, 1)
fp = np.sum(kn.stats[npos:] >= tr)
fdr += fp / cp
power += np.mean(kn.stats[0:npos] >= tr)
estimated_fdr = (np.sum(kn.stats <= -tr) /
(1 + np.sum(kn.stats >= tr)))
assert_array_equal(estimated_fdr < target_fdr, True)
power /= nrep
fdr /= nrep
assert_array_equal(power > 0.6, True)
assert_array_equal(fdr < target_fdr + 0.05, True)
import pandas as pd
import statsmodels.api as sm
import numpy as np
import patsy
import statsmodels.stats.knockoff_regeffects as kr
from nhanes_data import dx
dy = dx.copy()
dy.DMDEDUC2 = dy.DMDEDUC2.replace({7: np.nan, 9: np.nan})
dy.RIDAGEYR -= dy.RIDAGEYR.mean()
dy.BMXBMI -= dy.BMXBMI.mean()
dy.RIAGENDR = (dy.RIAGENDR == 2).astype(np.float64)
dy["SomeCollege"] = 1 * (dy.DMDEDUC2 >= 4)
yvec, xmat = patsy.dmatrices(
"BPXSY1 ~ 0 + RIDAGEYR * RIAGENDR * BMXBMI * SomeCollege",
data=dy,
return_type="dataframe")
yvec = yvec.values[:, 0]
yvec -= yvec.mean()
yvec /= yvec.std()
xcols = xmat.columns.tolist()
xmat = np.asarray(xmat)
xmat -= xmat.mean(0)
xmat /= xmat.std(0)
ko = sm.stats.RegressionFDR(yvec, xmat, kr.ForwardEffects(pursuit=True))
exoga = np.concatenate((exog1, exog2), axis=1)
gmat = np.dot(exoga.T, exoga)
cm1 = gmat[0:4, 0:4]
cm2 = gmat[4:, 4:]
cm3 = gmat[0:4, 4:]
assert_allclose(cm1, cm2, rtol=1e-4, atol=1e-4)
assert_allclose(cm1 - cm3, np.diag(sl * np.ones(4)), rtol=1e-5, atol=1e-5)
@pytest.mark.parametrize("p", [49, 50])
@pytest.mark.parametrize("tester", [
kr.CorrelationEffects(),
kr.ForwardEffects(pursuit=False),
kr.ForwardEffects(pursuit=True),
kr.OLSEffects(),
kr.RegModelEffects(sm.OLS),
kr.RegModelEffects(sm.OLS, True, fit_kws={
"L1_wt": 0,
"alpha": 1
}),
])
@pytest.mark.parametrize("method", ["equi", "sdp"])
def test_testers(p, tester, method):
if method == "sdp" and not has_cvxopt:
return
np.random.seed(2432)