Handling missing alternatives (NaNs) in multinomial logit with PyMC?

1

Hello,

I’m trying to build a simple multinomial logit model in PyMC using pm.Categorical + softmax.

My data is something like distances to different options, but not all options exist for every observation. So when I pivot to a matrix (obs × alternatives), I get NaNs.

Example:

walk = pm.Data(
    "walk",
    df.pivot(index="obs_id", columns="alt", values="distance").values,
    dims=["obs", "alt"]
)

This fails because of NaNs:

Masked arrays or arrays with nan entries are not supported

My question is:

What is the correct way in PyMC to handle alternatives that are not available for some observations?

Should I fill NaNs with large values (big distance)?

Or is there a proper way to define an availability mask in the model?

I want the model to understand that some alternatives are not part of the choice set, not just “bad”.

Any guidance or examples would be really helpful.

2

You can set the respective probabilities to -inf, which after the softmax are 0. As long as you don’t have any rows where all entries are impossible, it should be fine.

2 Likes
3

Really appreciate it.

4

Note also that @ricardoV94’s solution respects Luce’s choice axiom:

3 Likes