`@fitmodel` errors on sanction dataset with seed 123.

[codetalker7]

Consider the following code.

using RDatasets, CRRao

sanction = dataset("Zelig", "sanction")
CRRao_seed(123)

model = @fitmodel((Num ~ Target + Coop + NCost), sanction, NegBinomRegression(), Prior_Cauchy())

The above code returns an error (the error message is too long to print here).

Running the same code with a different seed (42) runs successfully, but with a warning (the image of the warning is attached).

[sourish-cmi]

This warning is not due to seed choice. This warning is due to the MCMC chain exploring some possible values and getting some illegal values at the initial stage. This warning is coming from Turing. Can't tell why it stops due to seed 123. Can you please share the error message from seed 123?

[ayushpatnaikgit]

NegativeBinomial: the condition zero(p) < p <= one(p) is not satisfied.

This is the error when we use seed = 123

[sourish-cmi]

This is very odd. This issue is supposed to be mathematically taken care of in the modeling. What is the probability that 1 is turning out as 1.000000002 ! Do you think it is possible?

[ayushpatnaikgit]

We should file this issue in Turing.jl

[sourish-cmi]

Let me prepare a note with code easy to understand for the Turing developer.

[ayushpatnaikgit]

This is the underlying code:

using Turing, Random, RDatasets, StatsModels
Random.seed!(123)
data = dataset("Zelig", "sanction")
formula = @formula(Num ~ Target + Coop + NCost)
formula = apply_schema(formula, schema(formula, data));
y, X = modelcols(formula, data);
@model NegBinomReg(X, y) = begin
    p = size(X, 2);
    n = size(X, 1);
    #priors
    λ~InverseGamma(1.0,1.0)
    ν~InverseGamma(1.0,1.0)
    α ~ TDist(ν)*λ
    β ~ filldist(TDist(ν)*λ, p)  
  
    ## link
    z = α .+ X * β
    mu = exp.(z)
    
    #likelihood
    for i = 1:n
        y[i] ~ NegativeBinomial(λ, (1 / (1 + mu[i]/ λ)))
    end
end
NegBinomReg_model=NegBinomReg(X,y);
chain = sample(NegBinomReg_model, NUTS(), 10000)
summaries, quantiles = describe(chain)

@sourish-cmi it is running perfectly. We aren't able to isolate the bug to report it on Turing.

[ayushpatnaikgit]

@sourish-cmi can you make a small code reproducible code for this for the Turing folks.