def p_tpr_fdr(p_values, intervals, truth, fdr_threshold=0.1):
interval_tpr = np.full((intervals.shape[0], intervals.shape[0]), np.nan)
interval_fdr = np.full(interval_tpr.shape, np.nan)
for idx1, interval in enumerate(intervals):
for idx2, interval in enumerate(intervals):
p = p_values[:, idx1, idx2]
if np.any(np.isnan(p)):
continue
p_indices = ~np.isnan(p)
p = p[p_indices]
t = truth[p_indices]
d = bh_predictions(p, fdr_threshold)
interval_tpr[idx1, idx2] = tpr(t, d)
interval_fdr[idx1, idx2] = fdr(t, d)
return interval_tpr, interval_fdr
importance = rf_importance(model.models)
else:
importance = linear_model_importance(model.models)
# Get the p-values and add correction term
all_p_filename = "data/{}/{}.npy".format(trial, info.prefix)
p_values[info.name][trial] = np.load(all_p_filename)
p_values[info.name][trial] = (p_values[info.name][trial] * nperms +
1) / (nperms + 1)
# Run the filtered testing procedure
important = importance >= importance_threshold
p_important = p_values[info.name][trial][important]
print("\tFiltering down to {} features".format(important.sum()))
pred = np.zeros(P, dtype=int)
pred[important] = bh_predictions(p_important, fdr_threshold)
tpr_vals[info.name][trial] = tpr(truth, pred)
fdr_vals[info.name][trial] = fdr(truth, pred)
labels = [info.name for info in infos]
print("Plotting power and FDR results")
results_plot(
[tpr_vals[label] for label in labels],
[fdr_vals[label] for label in labels],
[label.replace(" ", "\n") for label in labels],
fdr_threshold,
)
plt.savefig("plots/predictors-importance.pdf", bbox_inches="tight")
plt.close()
def add_if_finished(trial, all_p_values, all_tpr, all_fdr, truth, fdr_threshold):
if not np.any(np.isnan(all_p_values[trial])):
predictions = bh_predictions(all_p_values[trial], fdr_threshold)
all_tpr[trial] = tpr(truth, predictions)
all_fdr[trial] = fdr(truth, predictions)
):
p = np.load(p_filename)
p_values[info.name][trial, feature] = p
np.save(all_p_filename, p_values[info.name][trial])
p_values[info.name][trial] = (p_values[info.name][trial] * nperms + 1) / (
nperms + 1
)
missing = np.isnan(p_values[info.name][trial])
signal_p_values[info.name].extend(
p_values[info.name][trial, :S][~missing[:S]]
)
p_trial = np.ones(P)
p_trial[~missing] = p_values[info.name][trial][~missing]
pred = bh_predictions(p_trial, fdr_threshold)
tpr_vals[info.name][trial] = tpr(truth, pred)
fdr_vals[info.name][trial] = fdr(truth, pred)
if np.any(missing):
if missing.sum() > 10:
print("Total missing: {}".format(missing.sum()))
else:
print("Missing: {}".format(np.arange(P)[missing]))
# Load the p-values for the other models
for path, name in [
("cv_robust_nonlinear_p_values", "Neural Net"),
("cv_robust_linear_p_values", "OLS"),
]:
p_trial = np.load("data/{}/{}.npy".format(trial, path))
p_trial = (p_trial * nperms + 1) / (nperms + 1)
signal_p_values[name].extend(p_trial[:S])
def main():
N = 500 # total number of samples
P = 500 # number of features
S = 40 # number of signal features
nperms = 5000
nbootstraps = 100
fdr_threshold = 0.1
ntrials = 100
names = ['Holdout\nPermutation', 'Calibrated\nHRT\n(linear)', 'Uncalibrated\nHRT',
'CV\nPermutation', 'Calibrated\nCV-HRT\n(linear)', 'Uncalibrated\nCV-HRT',
'Calibrated\nHRT\n(linear)', 'Calibrated\nHRT',
'Calibrated\nCV-HRT\n(linear)', 'Calibrated\nCV-HRT']
prefixes = ['perm', 'linear', 'nonlinear',
'cv_perm', 'cv_linear', 'cv_nonlinear',
'robust_linear', 'robust_nonlinear',
'cv_robust_linear', 'cv_robust_nonlinear']
p_values = np.full((len(prefixes), ntrials, P), np.nan)
tpr_vals, fdr_vals = np.full((len(prefixes), ntrials), np.nan), np.full((len(prefixes), ntrials), np.nan)
for idx, prefix in enumerate(prefixes):
for trial in range(ntrials):
if (trial % 25) == 0:
print('{} trial: {}'.format(prefix, trial))
TRUTH_PATH = 'data/{}/truth.csv'.format(trial)
truth = np.loadtxt(TRUTH_PATH, delimiter=',')
P_VALUE_PATH = 'data/{}/{}_p_values.npy'.format(trial, prefix)
if os.path.exists(P_VALUE_PATH):
p_values[idx, trial] = np.load(P_VALUE_PATH)
clean_up_needed = False
if np.any(np.isnan(p_values[idx,trial])):
clean_up_needed = True
for feature in range(P):
Pi_PATH = 'data/{}/{}_p_values_{}.npy'.format(trial, prefix, feature)
if np.isnan(p_values[idx, trial, feature]) and os.path.exists(Pi_PATH):
try:
p_values[idx,trial,feature] = np.load(Pi_PATH)
except:
os.remove(Pi_PATH)
# p_values[idx, trial] = p_values[idx, trial] * nperms / (nperms+1)
missing = np.isnan(p_values[idx, trial])
pred = bh_predictions(p_values[idx, trial][~missing], fdr_threshold)
tpr_vals[idx, trial] = tpr(truth[~missing], pred)
fdr_vals[idx, trial] = fdr(truth[~missing], pred)
if not np.any(np.isnan(p_values[idx,trial])):
# clean up
if clean_up_needed:
np.save(P_VALUE_PATH, p_values[idx,trial])
for feature in range(P):
Pi_PATH = 'data/{}/{}_p_values_{}.npy'.format(trial, prefix, feature)
if os.path.exists(Pi_PATH):
# print('Would delete {}'.format((idx, trial, feature)))
os.remove(Pi_PATH)
else:
print('Trial {} Nulls: {}'.format(trial, np.where(np.isnan(p_values[idx, trial]))[0]))
if 'robust' in prefix:
# Get the distribution of confidence intervals
bounds = np.full((ntrials, P, 2), np.nan)
for trial in range(ntrials):
BOUNDS_PATH = 'data/{}/{}_bounds.npy'.format(trial, prefix)
if os.path.exists(BOUNDS_PATH):
bounds[trial] = np.load(BOUNDS_PATH)
clean_up_needed = False
if np.any(np.isnan(bounds[trial])):
clean_up_needed = True
for feature in range(P):
BOUNDS_i_PATH = 'data/{}/{}_bounds_{}.npy'.format(trial, prefix, feature)
if np.any(np.isnan(bounds[trial, feature])) and os.path.exists(BOUNDS_i_PATH):
bounds[trial,feature] = np.load(BOUNDS_i_PATH)
if not np.any(np.isnan(bounds[trial])):
# clean up
if clean_up_needed:
np.save(BOUNDS_PATH, bounds[trial])
for feature in range(P):
BOUNDS_i_PATH = 'data/{}/{}_bounds_{}.npy'.format(trial, prefix, feature)
if os.path.exists(BOUNDS_i_PATH):
# print('Would delete {}'.format(BOUNDS_i_PATH))
os.remove(BOUNDS_i_PATH)
bounds_plot(bounds)
plt.savefig('plots/liang-bounds-{}.pdf'.format(prefix.replace('_', '-')), bbox_inches='tight')
print('*** {} model ({} trials) ***'.format(names[idx], (~np.isnan(tpr_vals[idx])).sum()))
print('TPR: {:.2f}%'.format(np.nanmean(tpr_vals[idx], axis=0)*100))
print('FDR: {:.2f}%'.format(np.nanmean(fdr_vals[idx], axis=0)*100))
print('')
p_plot(p_values[idx], S)
plt.savefig('plots/liang-p-values-{}.pdf'.format(prefix.replace('_','-')), bbox_inches='tight')
selected = np.array([0,3,2,5,7,9])
results_plot([tpr_vals[i] for i in selected],
[fdr_vals[i] for i in selected],
[names[i] for i in selected],
fdr_threshold)
plt.savefig('plots/liang-tpr-fdr.pdf', bbox_inches='tight')
