[CLOSED] Help sharing dimension between location and scale in a simple hierarchical model

Questions version agnostic
, ,
1

Hello,

first of all, congratulations on the new and shiny version 4!

I’m trying to define a simple hierarchical model, and have a problem in understanding how to use implicit dimensions. My (simplified) code is here:

coords = {
  "dim1": ["one", "two", "three"],
  "dim2": list(range(50)),
}

with pm.Model(coords=coords) as m:
  # Define priors
  γ = pm.Normal("γ", -2.8, 0.3, dims=("dim1",))
  τ = pm.HalfNormal("τ", 1, dims=("dim1",))
  α = pm.Normal("α", γ, τ, dims=("dim2", "dim1"))
  
  # Define the log-link function
  linear_part = α[dim2_idx, dim1_idx] + at.log(exposure)
  μ = at.exp(linear_part)  
  
  # The likelihood
  y = pm.Poisson("y", μ, observed=claims)

When I try to sample from the prior I get an error IndexError: index 22 is out of bounds for axis 0 with size 3 and I know this is related to having both γ and τ as hyperpriors in the definition of α. This code works if I replace τ with a constant, so I suspect this has something to do with broadcasting, but I’m not really sure how to communicate to pymc that I want both γ and τ to be indexed by the same dim1_idx in the linear part of the GLM.

Here dim1_idx and dim2_idx are coming from categorical columns in a pandas DataFrame, and exposure and claims are pandas series with the same length as dim1_idx and dim2_idx.

Any help would be greatly appreciated!
Cheers,
Omri

2

OK, I realized I actuallly had a bug somewhere else in the code, the snippet I shared actually works!

Sorry for the trouble :slight_smile:

2 Likes