Survival analysis in PyMC 3

问题

I tried to port simple survival model from here (the first one in introduction) form PyMC 2 to PyMC 3. However, I didn't find any equivalent to "observed" decorator and my attempt to write a new distribution failed. Could someone provide an example how is this done in PyMC 3?

回答1:

This is a tricky port, and requires three new concepts:

1. Use of the theano tensor
2. Use of the DensityDist
3. Passing a dict as observed

This code provides the equivalent model as the PyMC2 version you linked to above:

import pymc3 as pm
from pymc.examples import melanoma_data as data
import theano.tensor as t

times = data.t # not to be confused with the theano tensor t!
failure = (data.censored==0).astype(int)

with pm.Model() as model:

    beta0 = pm.Normal('beta0', mu=0.0, tau=0.0001)
    beta1 = pm.Normal('beta1', mu=0.0, tau=0.0001)
    lam = t.exp(beta0 + beta1*data.treat)

    def survival_like(failure, value):
        return t.sum(failure * t.log(lam) - lam * value)

    survive = pm.DensityDist('survive', survival_like,
                        observed={'failure': failure, 'value': times})

with model:

    start = pm.find_MAP()
    step = pm.NUTS(scaling=start)
    trace = pm.sample(10000, step=step, start=start)

pm.traceplot(trace);

Output as follows:

来源：https://stackoverflow.com/questions/22015055/survival-analysis-in-pymc-3