Using Bound RV in a Mixture

Questions v5
,
1

I am trying to define a mixture of bounded random variables e.g.

with pm.Model() as model:
    w =  w = pm.Dirichlet("w", a=np.ones(3))
    upper_sig = pm.HalfNormal("upper_sig", sigma=0.2)
    lower_sig = pm.HalfNormal("lower_sig", sigma=0.2)

    efficiencies = pm.Mixture(
        "efficiencies",
        w=w,
        comp_dists=[
            pm.Bound(pm.Normal, lower=1.0).dist(mu=1.0, sigma=upper_sig),
            pm.Bound(pm.Normal, lower=0.0, upper=1.0).dist(mu=1.0, sigma=lower_sig),
            pm.Uniform.dist(lower=0.0, upper=1.0),
        ],
        observed=data,
    )

but I am getting the following error:

TypeError: __new__() missing 1 required positional argument: 'dist'

Issue seems related to Using a bounded variable within a Mixture - Questions - PyMC Discourse - however, in the discussion it was mentioned the approach wouldn’t work for latest version of the code.

2

Ok, looks like the dists need to be constructed like this:

...
comp_dists=[
    pm.Bound.dist(pm.Normal.dist(mu=1.0, sigma=upper_sig), lower=1.0),
    pm.Bound.dist(pm.Normal.dist(mu=1.0, sigma=lower_sig), lower=0.0, upper=1.0),
...
3

I would suggest you use pm.Truncated instead. That way you can also perform prior and posterior predictive sampling with the Mixture model.

1 Like
4

Perfect, I could see Bound was broken for the prior and posterior sampling, but that Censored was broken for sampling! Truncated works as you suggest.

5

Broken is perhaps a strong word :slight_smile: They are different objects

Anyway glad I could help

1 Like
6

Apologies, totally agree - just my naive understanding of the classes! :slight_smile:

1 Like