def do_km(name, time, censor, split, outdir):
"""Given three clean (pre-processed) lists, make a kmplot of the data, and save it to outdir"""
data = {
'time': robjects.IntVector(np.array(time)),
'censor': robjects.IntVector(np.array(censor)),
'split': robjects.IntVector(np.array(split))
}
df = robjects.DataFrame(data)
surv = importr('survival')
grdevices = importr('grDevices')
km = surv.survfit(robjects.Formula('Surv(time, censor) ~ split'), data=df)
grdevices.png(file=os.path.join(outdir, name + '_km.png'),
width=512,
height=512)
r.plot(km,
xlab='Time',
ylab='Cumulative Hazard',
col=robjects.StrVector(['Red', 'Blue']))
r.legend(1000,
1,
robjects.StrVector(['<= Mean', '> Mean']),
lty=robjects.IntVector([1, 1]),
col=robjects.StrVector(['Red', 'Blue']))
grdevices.dev_off()
def val(self):
""" Estimate value functions with b-splines and compare """
new_data = pd.DataFrame({'OverallRank': np.linspace(1, 194, 1000)})
fit_a = self.spline_est(self.policy_a['value'], new_data)
fit_b = self.spline_est(self.policy_b['value'], new_data)
r.pdf(os.path.join(os.path.dirname(self.out_dir), 'value.pdf'))
r.plot(new_data['OverallRank'], fit_a, type='l', xlab='Rank_M',
ylab='V(Rank)')
r.lines(new_data['OverallRank'], fit_b, col='red')
r.points(self.policy_a['value']['OverallRank'],
self.policy_a['value']['val'],
col='black')
r.points(self.policy_b['value']['OverallRank'],
self.policy_b['value']['val'],
col='red')
r.legend('topright', np.array(['No Info', 'Info']),
lty=np.array([1, 1]), col=np.array(['black', 'red']))
r('dev.off()')
diff = np.array(fit_b) - np.array(fit_a)
r.pdf(os.path.join(os.path.dirname(self.out_dir), 'value_diff.pdf'))
r.plot(new_data['OverallRank'], diff, type='l', xlab='Rank',
ylab='V(Rank|info=1) - V(Rank|info=0)')
r.abline(h=0, lty=2)
r('dev.off()')
diff = (np.array(fit_b) - np.array(fit_a)) / np.array(fit_a)
r.pdf(os.path.join(os.path.dirname(self.out_dir),
'value_percent_diff.pdf'))
r.plot(new_data['OverallRank'], diff, type='l', xlab='Rank',
ylab='(V(Rank|info=1) - V(Rank|info=0)) / V(Rank|info=0)')
r.abline(h=0, lty=2)
r('dev.off()')
data_path = dirname(dirname(__file__))
data_path = join(data_path, 'data', 'lawData.csv')
data = pd.read_csv(data_path)
new_data = deepcopy(data.loc[data['year'] == 2013, 'OverallRank'])
#new_data = np.concatenate((
# new_data, np.zeros(lc.N_SCHOOLS - len(new_data))
#))
new_data = pd.DataFrame({'OverallRank': np.array(new_data)})
fit_a = self.spline_est(self.policy_a['value'], new_data)
fit_b = self.spline_est(self.policy_b['value'], new_data)
diff = np.sum(np.array(fit_b) - np.array(fit_a))
pdiff = diff / np.sum(fit_a)
print(" - Change in Producer Surplus: {0}".format(diff))
print(" - Percent change in Producer Surplus: {0}".format(pdiff))
return diff
def draw_survival_curves(feature,
surv,
assignment=None,
filename='tmp.png',
show=False,
title=True,
labels=None,
colors=['blue', 'red'],
ann=None,
show_legend=True,
q=.25,
std=None):
if assignment is None:
num_panels = 1
assignment = feature.map(lambda s: 1)
name = lambda v: str(feature.name) if feature.name != None else ''
else:
num_panels = len(assignment.unique())
name = lambda v: str(assignment.name) + ' = ' + str(v)
if (labels is None) and ((len(feature) / feature.nunique()) > 10):
labels = r.sort(r.c(*feature.unique())) # R sorts bad
colors = ['blue', 'green', 'red', 'cyan', 'magenta', 'yellow', 'black']
if feature.dtype == 'bool':
feature = feature.map({True: 'True', False: 'False'})
r.png(filename=filename, width=200 * (num_panels + 1), height=300, res=75)
fmla = robjects.Formula('Surv(days, event) ~ feature')
r.par(mfrow=r.c(1, num_panels))
r.par(mar=r.c(4, 5, 4, 1))
r.par(xpd=True)
if (get_vec_type(feature) == 'real') and (len(feature.unique()) > 10):
colors = ['blue', 'orange', 'red']
if q == .5:
labels = ['Bottom 50%', 'Top 50%']
else:
labels = [
'Bottom {}%'.format(int(q * 100)), 'Normal',
'Top {}%'.format(int(q * 100))
]
ls = r.c(*colors)
def plot_me(sub_f, label):
if (get_vec_type(sub_f) == 'real') and (len(sub_f.unique()) > 10):
sub_f = to_quants(sub_f, q=q, std=std)
m = get_cox_ph(surv, sub_f, formula=fmla)
r_data = m.rx2('call')[2]
p = log_rank(sub_f, surv)['p']
ls = r.c(*colors)
r.plot(survival.survfit(fmla, r_data),
lty=1,
col=ls,
lwd=4,
cex=1.25,
xlab='Years to Event',
ylab='Survival')
r.title(label, cex=3.)
if ann == 'p':
r.text(.2, 0, labels='logrank p = {0:.1e}'.format(p), pos=4)
elif ann != None:
r.text(0, labels=ann, pos=4)
if show_legend == 'out':
r.par(xpd=True, mar=r.c(4, 5, 5, 8))
for value in sorted(assignment.ix[feature.index].dropna().unique()):
f = feature.ix[assignment[assignment == value].index]
if len(f.unique()) > 1:
plot_me(f, name(value))
if show_legend == True:
mean_s = surv.ix[:, 'event'].ix[assignment[assignment ==
value].index].mean()
if mean_s < .5:
r.legend(surv.ix[:, 'days'].max() * .05 / 365.,
.45,
labels,
lty=1,
col=ls,
lwd=3,
bty='o')
else:
r.legend(surv.ix[:, 'days'].max() * .4 / 365,
.9,
labels,
lty=1,
col=ls,
lwd=3,
bty='o')
elif show_legend == 'out':
r.legend(surv.ix[:, 'days'].max() * 1.1 / 365,
.9,
labels,
lty=1,
col=ls,
lwd=3,
bty='o')
r('dev.off()')
if show:
return Show(filename)
def draw_survival_curves(feature, surv, assignment=None, filename='tmp.png', show=False,
title=True, labels=None, colors=['blue', 'red'], ann=None,
show_legend=True, q=.25, std=None):
if assignment is None:
num_panels = 1
assignment = feature.map(lambda s: 1)
name = lambda v: str(feature.name) if feature.name != None else ''
else:
num_panels = len(assignment.unique())
name = lambda v: str(assignment.name) + ' = ' + str(v)
if (labels is None) and ((len(feature) / feature.nunique()) > 10):
labels = r.sort(r.c(*feature.unique())) # R sorts bad
colors = ['blue', 'green', 'red', 'cyan', 'magenta', 'yellow', 'black']
if feature.dtype == 'bool':
feature = feature.map({True: 'True', False: 'False'})
r.png(filename=filename, width=200 * (num_panels + 1), height=300, res=75)
fmla = robjects.Formula('Surv(days, event) ~ feature')
r.par(mfrow=r.c(1, num_panels))
r.par(mar=r.c(4, 5, 4, 1))
r.par(xpd=True)
if (get_vec_type(feature) == 'real') and (len(feature.unique()) > 10):
colors = ['blue', 'orange', 'red']
if q == .5:
labels = ['Bottom 50%', 'Top 50%']
else:
labels = ['Bottom {}%'.format(int(q * 100)), 'Normal', 'Top {}%'.format(int(q * 100))]
ls = r.c(*colors)
def plot_me(sub_f, label):
if (get_vec_type(sub_f) == 'real') and (len(sub_f.unique()) > 10):
sub_f = to_quants(sub_f, q=q, std=std)
m = get_cox_ph(surv, sub_f, formula=fmla)
r_data = m.rx2('call')[2]
p = log_rank(sub_f, surv)['p']
ls = r.c(*colors)
r.plot(survival.survfit(fmla, r_data), lty=1, col=ls, lwd=4, cex=1.25,
xlab='Years to Event', ylab='Survival');
r.title(label, cex=3.)
if ann == 'p':
r.text(.2, 0, labels='logrank p = {0:.1e}'.format(p), pos=4)
elif ann != None:
r.text(0, labels=ann, pos=4)
if show_legend == 'out':
r.par(xpd=True, mar=r.c(4, 5, 5, 8))
for value in sorted(assignment.ix[feature.index].dropna().unique()):
f = feature.ix[assignment[assignment == value].index]
if len(f.unique()) > 1:
plot_me(f, name(value))
if show_legend == True:
mean_s = surv.ix[:, 'event'].ix[assignment[assignment == value].index].mean()
if mean_s < .5:
r.legend(surv.ix[:, 'days'].max() * .05 / 365., .45, labels,
lty=1, col=ls, lwd=3, bty='o')
else:
r.legend(surv.ix[:, 'days'].max() * .4 / 365, .9, labels,
lty=1, col=ls, lwd=3, bty='o')
elif show_legend == 'out':
r.legend(surv.ix[:, 'days'].max() * 1.1 / 365, .9, labels,
lty=1, col=ls, lwd=3, bty='o')
r('dev.off()')
if show:
return Show(filename)