def test_gaussian(n=200, p=30, burnin=2000, ndraw=8000,
compute_interval=False,
s=6):
X, y, beta, lasso_active, sigma = lasso_instance(n=n,
p=p)
n, p = X.shape
lam = sigma * np.mean(np.fabs(np.dot(X.T, np.random.standard_normal((n, 10000)))).max(0))
L = randomized_lasso(y, X, lam, (True, 0.4 * np.diag(np.sqrt(np.diag(np.dot(X.T, X))))),
sandwich=False)
L.fit()
L = randomized_lasso(y, X, lam, (True, 0.4 * np.diag(np.sqrt(np.diag(np.dot(X.T, X))))),
sandwich=True)
L.fit()
if (set(range(s)).issubset(L.active) and
L.active.shape[0] > s):
L.constraints.mean[:p] = 0 * L.unbiased_estimate
v = np.zeros_like(L.active)
v[s] = 1.
P0, interval = L.hypothesis_test(v, burnin=burnin, ndraw=ndraw,
compute_interval=compute_interval)
target = (beta[L.active]*v).sum()
estimate = (L.unbiased_estimate[:L.active.shape[0]]*v).sum()
low, hi = interval
v = np.zeros_like(L.active)
v[0] = 1.
PA, _ = L.hypothesis_test(v, burnin=burnin, ndraw=ndraw,
compute_interval=compute_interval)
return P0, PA, L
def test_gaussian(n=200,
p=30,
burnin=2000,
ndraw=8000,
compute_interval=False,
sandwich=True,
selected=False,
s=6,
snr=7,
nsim=None):
X, y, beta, lasso_active, sigma = lasso_instance(n=n,
p=p,
snr=snr,
s=s,
rho=0.1)
n, p = X.shape
lam = 2. * sigma * np.mean(
np.fabs(np.dot(X.T, np.random.standard_normal((n, 10000)))).max(0))
L = randomized_lasso(
y,
X,
lam, (True, 0.8 * sigma * np.diag(np.sqrt(np.diag(np.dot(X.T, X))))),
sandwich=sandwich,
selected=selected,
dispersion=sigma**2)
L.fit()
if (set(range(s)).issubset(L.active) and L.active.shape[0] > s):
L.unbiased_estimate[:] = np.zeros(p)
L.constraints.mean[:p] = 0 * L.unbiased_estimate
v = np.zeros_like(L.active)
v[s] = 1.
P0, interval = L.hypothesis_test(v,
burnin=burnin,
ndraw=ndraw,
compute_interval=compute_interval)
target = (beta[L.active] * v).sum()
estimate = (L.unbiased_estimate[:L.active.shape[0]] * v).sum()
low, hi = interval
v = np.zeros_like(L.active)
v[0] = 1.
PA, _ = L.hypothesis_test(v,
burnin=burnin,
ndraw=ndraw,
compute_interval=compute_interval)
return P0, PA, L
def test_gaussian(n=200, p=30, burnin=2000, ndraw=8000,
compute_interval=False,
sandwich=True,
selected=False,
s=6,
snr=7,
condition_on_more=False,
nsim=None):
X, y, beta, lasso_active, sigma = lasso_instance(n=n,
p=p,
snr=snr,
s=s,
rho=0.1)
n, p = X.shape
lam = 2. * sigma * np.mean(np.fabs(np.dot(X.T, np.random.standard_normal((n, 10000)))).max(0))
L = randomized_lasso(y, X, lam, (True, 0.8 * sigma * np.diag(np.sqrt(np.diag(np.dot(X.T, X))))),
sandwich=sandwich,
selected=selected,
dispersion=sigma**2)
L.fit()
if (set(range(s)).issubset(L.active) and
L.active.shape[0] > s):
L.unbiased_estimate = np.zeros(p)
L.constraints.mean[:p] = 0 * L.unbiased_estimate
v = np.zeros_like(L.active)
v[s] = 1.
P0, interval = L.hypothesis_test(v, burnin=burnin, ndraw=ndraw,
compute_interval=compute_interval,
condition_on_more=condition_on_more)
target = (beta[L.active]*v).sum()
estimate = (L.unbiased_estimate[:L.active.shape[0]]*v).sum()
low, hi = interval
v = np.zeros_like(L.active)
v[0] = 1.
PA, _ = L.hypothesis_test(v, burnin=burnin, ndraw=ndraw,
compute_interval=compute_interval,
condition_on_more=condition_on_more)
return P0, PA, L