pymc.dims.Censored#

class pymc.dims.Censored(name, *dist_params, dims=None, initval=None, observed=None, total_size=None, transform=UNSET, default_transform=UNSET, model=None, **kwargs)[source]#

Censored distribution.

The pdf of a censored distribution is

\[\begin{split}\begin{cases} 0 & \text{for } x < lower, \\ \text{CDF}(lower, dist) & \text{for } x = lower, \\ \text{PDF}(x, dist) & \text{for } lower < x < upper, \\ 1-\text{CDF}(upper, dist) & \text {for} x = upper, \\ 0 & \text{for } x > upper, \end{cases}\end{split}\]
Parameters:
distunnamed_distribution

Univariate distribution which will be censored. This distribution must have a logcdf method implemented for sampling.

Warning

dist will be cloned, rendering it independent of the one passed as input.

lowerxtensor_like of float, optional

Lower (left) censoring point, broadcast against dist. An entry of -np.inf leaves that entry uncensored. Default is -np.inf.

upperxtensor_like of float, optional

Upper (right) censoring point, broadcast against dist. An entry of np.inf leaves that entry uncensored. Default is np.inf.

Warning

Continuous censored distributions should only be used as likelihoods. Continuous censored distributions are a form of discrete-continuous mixture and as such cannot be sampled properly without a custom step sampler. If you wish to sample such a distribution, you can add the latent uncensored distribution to the model and then wrap it in a Deterministic clip().

Examples

with pm.Model():
    normal_dist = pm.Normal.dist(mu=0.0, sigma=1.0)
    censored_normal = pm.Censored("censored_normal", normal_dist, lower=-1, upper=1)

Censoring bounds can differ per entry:

with pm.Model():
    normal_dist = pm.Normal.dist(mu=0.0, sigma=1.0, shape=2)
    censored_normal = pm.Censored(
        "censored_normal", normal_dist, lower=[-1.0, -np.inf], upper=[1.0, 0.0]
    )

Methods

Censored.dist(dist, *[, lower, upper])