Best practices for dealing with NaNs in multiple data set using Turing

Specific Domains Probabilistic programming
, ,
1

What is the best way to deal with NaNs in multiple data treatments when using Turing?

For example, for this example tutorial from https://github.com/TuringLang/TuringTutorials/blob/master/tutorials/10-bayesian-differential-equations/10_bayesian-differential-equations.jmd:

@model function fitlv(data, prob)
    σ ~ InverseGamma(2,3)
    α ~ truncated(Normal(1.3,0.5),0.5,2.5)
    β ~ truncated(Normal(1.2,0.25),0.5,2)
    γ ~ truncated(Normal(3.2,0.25),2.2,4.0)
    δ ~ truncated(Normal(1.2,0.25),0.5,2.0)
    ϕ1 ~ truncated(Normal(0.12,0.3),0.05,0.25)
    ϕ2 ~ truncated(Normal(0.12,0.3),0.05,0.25)
    p = [α,β,γ,δ,ϕ1,ϕ2]
    prob = remake(prob, p=p)
    predicted = solve(prob,SOSRI(),saveat=0.1)

    if predicted.retcode != :Success
        Turing.acclogp!(_varinfo, -Inf)
    end
    for j in 1:length(data)
        for i = 1:length(predicted)
            data[i,j] ~ MvNormal(predicted[i],σ)
        end
    end
end;

Here, how Turing is going to deal with NaN in data[i,j]? Assuming the number of treatments [j] is high, do you have any recommendations on how to keep it simple and deal with the nans?
Thanks!

2

I believe here Turing will error, because it doesn’t know what the NaNs mean. So, what do the NaNs mean?

3

In my case the NaNs mean missing. I didn’t tried using missing. How does turing is dealing with this?
I found another way to solve this problem by estimate the likelihood individually for each set of data and not using the MvNormal. But I imagine that is slowing down the process.