def plot(self, variable, label=None,
color=None, ref_val=None, alpha=.25,
bins=None, title=None,
hdi=None, outfile=None,
ref_color=None):
"""
Plot the histogram of a variable trace
Parameters
----------
variable : str
The name of one of self.variables to plot
label : str
Alternative label for the legend
ref_val : float
A reference value location at which to draw a vertical line
alpha : float in [0 1)
The transparency of the histogram
bins : int
The number of histogram bins
title : str
The title of the plot
hdi : float in [0, 1]
The amount of probability mass within the Highest Density Interval
to display on the histogram.
outfile : str
The name of an output file to save the figure to.
"""
from matplotlib import pyplot as plt # lazy import
from abra.vis import plot_interval
if variable not in self.variables:
print(self.variables)
raise ValueError('Variable `{}` not available'.format(variable))
label = label if label else variable
trace = getattr(self, variable)
if bins is None:
bins = int(len(trace.data) / 50.)
trace.hist(color=color, alpha=alpha, bins=bins, ref_val=ref_val, label=label)
if hdi is not None: # highest density interval
median = round(trace.percentiles(50), 3)
_hdi = [round(h, 3) for h in trace.hdi(1 - hdi)]
plot_interval(*_hdi, middle=median, display_text=True, color=color, offset=5)
if title is None:
if ref_val is not None:
gt = round(100 * trace.prob_greater_than(ref_val))
title = " {}% < {} = {} < {}%".format(100 - gt, variable, ref_val, gt)
else:
title = ''
plt.title(title, fontsize=16)
if outfile:
plt.savefig(outfile)
def visualize_rates_results(results,
figsize=(15, 10),
outfile=None,
*args,
**kwargs):
fig, axs = plt.subplots(3, 1, figsize=figsize)
# Sample Comparison plot
plt.sca(axs[0])
control_pmf = Poisson(results.control.mean,
color=CONTROL_COLOR,
label=results.control.name)
variation_pmf = Poisson(results.variation.mean,
color=VARIATION_COLOR,
label=results.variation.name)
control_pmf.plot(plot_type='bar', alpha=.5)
variation_pmf.plot(plot_type='bar', alpha=.5)
plt.legend()
plt.title("Sample Comparison")
# Rates +/- standard error plot
plt.sca(axs[1])
y_min, y_max = plt.ylim()
y_dist = (y_max - y_min) / LABEL_Y_OFFSET_FACTOR
plot_interval(*results.control.std_err(),
middle=results.control.mean,
y=y_dist,
offset=-0.015,
color=CONTROL_COLOR,
display_text=True,
label=results.control.name)
plot_interval(*results.variation.std_err(),
middle=results.variation.mean,
y=-y_dist,
offset=0.005,
color=VARIATION_COLOR,
display_text=True,
label=results.variation.name)
plt.legend()
plt.gca().get_yaxis().set_ticks([])
plt.title("Rates +/- Standard Error")
# Differences plot
plt.sca(axs[2])
plot_interval(*results.ci[0],
middle=results.delta,
color=DIFF_COLOR,
display_text=True)
plt.axvline(1., color=DIFF_COLOR, linestyle='--', linewidth=1.5)
plt.gca().get_yaxis().set_ticks([])
plt.title(results.comparison_type)
if outfile:
plt.savefig(outfile, bbox_inches='tight', dpi=300)
def visualize_bootstrap_results(results,
figsize=(15, 10),
outfile=None,
plot_type='bar',
*args,
**kwargs):
fig, axs = plt.subplots(3, 1, figsize=figsize)
# Sample Comparison plot
plt.sca(axs[0])
if plot_type == 'bar':
bins = 50 if results.control.nobs >= 100 or results.variation.nobs >= 100 else 20
results.control.hist(bins=bins,
color=CONTROL_COLOR,
alpha=.5,
label=results.control.name)
results.variation.hist(bins=bins,
color=VARIATION_COLOR,
alpha=.5,
label=results.variation.name)
else:
control_pmf = KdePdf(samples=results.control.data,
color=CONTROL_COLOR,
label=results.control.name)
variation_pmf = KdePdf(samples=results.variation.data,
color=VARIATION_COLOR,
label=results.variation.name)
control_pmf.plot(alpha=.5)
variation_pmf.plot(alpha=.5)
plt.legend()
plt.title("Sample Comparison")
# Bootstrapped statistic +/- HDI
plt.sca(axs[1])
y_min, y_max = plt.ylim()
y_dist = (y_max - y_min) / LABEL_Y_OFFSET_FACTOR
plot_interval(*results.aux['control'].hdi(),
middle=results.aux['control'].mean,
y=y_dist,
offset=-0.015,
color=CONTROL_COLOR,
display_text=True,
label=results.control.name)
plot_interval(*results.aux['variation'].hdi(),
middle=results.aux['variation'].mean,
y=-y_dist,
offset=0.005,
color=VARIATION_COLOR,
display_text=True,
label=results.variation.name)
plt.legend()
plt.gca().get_yaxis().set_ticks([])
plt.title(f"Bootstrap({results.test_statistic}) +/- 95% HDI")
# Differences plot
plt.sca(axs[2])
plot_interval(*results.ci[0],
middle=results.delta,
color=DIFF_COLOR,
display_text=True)
plt.axvline(0., color=DIFF_COLOR, linestyle='--', linewidth=1.5)
plt.gca().get_yaxis().set_ticks([])
plt.title(f"{results.comparison_type}({results.test_statistic})")
if outfile:
plt.savefig(outfile, bbox_inches='tight', dpi=300)
def visualize_binomial_results(results,
figsize=(15, 10),
outfile=None,
*args,
**kwargs):
"""
Visualize the results that use Gaussian approximation.
"""
tol = 1e-4
pmf_control = Binomial(p=results.control.mean,
n=results.control.nobs,
label=results.control.name,
color=CONTROL_COLOR)
pmf_variation = Binomial(p=results.variation.mean,
n=results.variation.nobs,
label=results.variation.name,
color=VARIATION_COLOR)
xy_control = zip(pmf_control.xgrid(),
pmf_control.density(pmf_control.xgrid()))
xy_variation = zip(pmf_variation.xgrid(),
pmf_variation.density(pmf_variation.xgrid()))
valid_xy_control = sorted([x for x in xy_control if x[1] >= tol],
key=lambda x: x[0])
valid_xy_variation = sorted([x for x in xy_variation if x[1] >= tol],
key=lambda x: x[0])
x_min = int(min(valid_xy_control[0][0], valid_xy_variation[0][0]))
x_max = int(max(valid_xy_control[-1][0], valid_xy_variation[-1][0]))
mean_diff = results.variation.mean - results.control.mean
std_diff = (results.control.var / results.control.nobs + \
results.variation.var / results.control.nobs) ** .5
pdf_diff = Gaussian(mean_diff,
std_diff,
label='Difference',
color=DIFF_COLOR)
fig, axs = plt.subplots(3, 1, figsize=figsize)
plt.sca(axs[0])
# make plotting more scalable
if pmf_control.n > 1000 or pmf_variation.n > 1000:
plot_type = 'step'
else:
plot_type = 'bar'
pmf_control.plot(plot_type=plot_type, alpha=.5)
pmf_variation.plot(plot_type=plot_type, alpha=.5)
raise_y(axs[0])
plt.xlim(x_min, x_max)
# plt.gca().get_xaxis().set_ticks([])
# plt.gca().get_yaxis().set_ticks([])
plt.legend()
plt.title("Sample Comparison")
plt.sca(axs[1])
y_min, y_max = plt.ylim()
y_dist = (y_max - y_min) / LABEL_Y_OFFSET_FACTOR
plot_interval(*results.control.std_err(),
middle=results.control.mean,
y=y_dist,
offset=-0.015,
color=CONTROL_COLOR,
display_text=True,
label=results.control.name)
plot_interval(*results.variation.std_err(),
middle=results.variation.mean,
y=-y_dist,
offset=0.005,
color=VARIATION_COLOR,
display_text=True,
label=results.variation.name)
plt.legend()
plt.gca().get_yaxis().set_ticks([])
plt.title("Proportions +/- Standard Error")
# Differences plot
plt.sca(axs[2])
plt.axvline(0., color=DIFF_COLOR, linestyle='--', linewidth=1.5)
# xs = pdf_diff.xgrid()
if results.inference_procedure.hypothesis == 'larger':
left_bound = results.ci[0][0]
right_bound = np.inf
elif results.inference_procedure.hypothesis == 'smaller':
right_bound = results.ci[0][1]
left_bound = np.inf
else:
left_bound = results.ci[0][0]
right_bound = results.ci[0][1]
plot_interval(left_bound,
right_bound,
mean_diff,
color=DIFF_COLOR,
display_text=True)
plt.gca().get_yaxis().set_ticks([])
plt.title(results.comparison_type)
if outfile:
plt.savefig(outfile, bbox_inches='tight', dpi=300)
def visualize_gaussian_results(results,
figsize=(15, 10),
outfile=None,
*args,
**kwargs):
"""
Visualize the results that use Gaussian approximation.
"""
pdf_control = Gaussian(mean=results.control.mean,
std=results.control.std,
label=results.control.name,
color=CONTROL_COLOR)
pdf_variation = Gaussian(mean=results.variation.mean,
std=results.variation.std,
label=results.variation.name,
color=VARIATION_COLOR)
pdfs = Pdfs([pdf_control, pdf_variation])
mean_diff = results.variation.mean - results.control.mean
std_diff = ((results.control.var / results.control.nobs) + \
(results.variation.var / results.control.nobs)) ** .5
pdf_diff = Gaussian(mean_diff,
std_diff,
label='Difference',
color=DIFF_COLOR)
fig, axs = plt.subplots(3, 1, figsize=figsize)
plt.sca(axs[0])
pdfs.plot()
raise_y(axs[0])
plt.gca().get_yaxis().set_ticks([])
plt.title("Sample Comparison")
x_min, x_max = plt.xlim()
plt.sca(axs[1])
y_min, y_max = plt.ylim()
y_dist = (y_max - y_min) / LABEL_Y_OFFSET_FACTOR
plot_interval(*results.control.std_err(),
middle=results.control.mean,
y=y_dist,
offset=-.015,
color=CONTROL_COLOR,
display_text=True,
label=results.control.name)
plot_interval(*results.variation.std_err(),
middle=results.variation.mean,
y=-y_dist,
offset=0.005,
color=VARIATION_COLOR,
display_text=True,
label=results.variation.name)
plt.legend()
plt.xlim(x_min, x_max)
plt.gca().get_yaxis().set_ticks([])
plt.title("Mean +/- Standard Error")
# plot differences distribution
plt.sca(axs[2])
plt.axvline(0., color=DIFF_COLOR, linestyle='--', linewidth=1.5)
if results.inference_procedure.hypothesis == 'larger':
left_bound = results.ci[0][0]
right_bound = np.inf
elif results.inference_procedure.hypothesis == 'smaller':
right_bound = results.ci[0][1]
left_bound = np.inf
else:
left_bound = results.ci[0][0]
right_bound = results.ci[0][1]
plot_interval(left_bound,
right_bound,
mean_diff,
color=DIFF_COLOR,
display_text=True)
plt.gca().get_yaxis().set_ticks([])
plt.title(results.comparison_type)
if outfile:
plt.savefig(outfile, bbox_inches='tight', dpi=300)