hammerlab/survivalstan

Name: survivalstan

Owner: Hammer Lab

Description: Library of Stan Models for Survival Analysis

Created: 2016-06-29 16:49:49.0

Updated: 2018-01-15 16:58:06.0

Pushed: 2017-10-24 20:08:51.0

Homepage:

Size: 6860

Language: Jupyter Notebook

GitHub Committers

User	Most Recent Commit	# Commits

Other Committers

User	Email	Most Recent Commit	# Commits

README

survivalstan: Survival Models in Stan

author: Jacki Novik

Overview

Library of Stan Models for Survival Analysis

Features:

• Variety of standard survival models
  • Weibull, Exponential, and Gamma parameterizations
  • PEM models with variety of baseline hazards
  • PEM model with varying-coefficients (by group)
  • PEM model with time-varying-effects
• Extensible framework - bring your own Stan code, or edit the models above
• Uses pandas data frames & patsy formulas
• Graphical posterior predictive checking (currently PEM models only)
• Plot posterior estimates of key parameters using seaborn
• Annotate posterior draws of parameter estimates, format as pandas dataframes
• Works with extensions to pystan, such as stancache or pystan-cache

Support

Documentation is available online.

For help, please reach out to us on gitter.

Installation / Usage

Install using pip, as:

$ pip install survivalstan

Or, you can clone the repo:

$ git clone https://github.com/hammerlab/survivalstan.git
$ pip install .

Contributing

Please contribute to survivalstan development by letting us know if you encounter any bugs or have specific feature requests.

In addition, we welcome contributions of:

• Stan code for survival models
• Worked examples, as jupyter notebooks or markdown documents

Usage examples

There are several examples included in the example-notebooks, roughly one corresponding to each model.

If you are not sure where to start, Test pem_survival_model with simulated data.ipynb contains the most explanatory text. Many of the other notebooks are sparse on explanation, but do illustrate variations on the different models.

For basic usage:

rt survivalstan
rt stanity
rt seaborn as sb
rt matplotlib.pyplot as plt
rt statsmodels

oad flchain test data from R's `survival` package
set = statsmodels.datasets.get_rdataset(package = 'survival', dataname = 'flchain' )
 dataset.data.query('futime > 7')
set_index(level = 0, inplace = True)

.g. fit Weibull survival model
fit_wei = survivalstan.fit_stan_survival_model(
model_cohort = 'Weibull model',
model_code = survivalstan.models.weibull_survival_model,
df = d,
time_col = 'futime',
event_col = 'death',
formula = 'age + sex',
iter = 3000,
chains = 4,
make_inits = survivalstan.make_weibull_survival_model_inits
)

oefplot for Weibull coefficient estimates
oxplot(x = 'value', y = 'variable', data = testfit_wei['coefs'])

r, use plot_coefs
ivalstan.utils.plot_coefs([testfit_wei])

rint summary of MCMC draws from posterior for each parameter
t(testfit_wei['fit'])


.g. fit Piecewise-exponential survival model 
g = survivalstan.prep_data_long_surv(d, time_col = 'futime', event_col = 'death')
fit_pem = survivalstan.fit_stan_survival_model(
model_cohort = 'PEM model',
model_code = survivalstan.models.pem_survival_model,
df = dlong,
sample_col = 'index',
timepoint_end_col = 'end_time',
event_col = 'end_failure',
formula = 'age + sex',
iter = 3000,
chains = 4,
)

rint summary of MCMC draws from posterior for each parameter
t(testfit_pem['fit'])

oefplot for PEM model results
oxplot(x = 'value', y = 'variable', data = testfit_pem['coefs'])

lot baseline hazard (only PEM models)
ivalstan.utils.plot_coefs([testfit_pem], element='baseline')

osterior-predictive checking (only PEM models)
ivalstan.utils.plot_pp_survival([testfit_pem])

.g. compare models using PSIS-LOO
ity.loo_compare(testfit_wei['loo'], testfit_pem['loo'])

ompare coefplots 
oxplot(x = 'value', y = 'variable', hue = 'model_cohort',
data = testfit_pem['coefs'].append(testfit_wei['coefs']))
legend(bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.)

or, use survivalstan.utils.plot_coefs)
ivalstan.utils.plot_coefs([testfit_wei, testfit_pem])