Experiment for Latent Gaussian Models

[avehtari]

@charlesm93

I think the most important thing is to figure out how the global parameter ϕ\phi drives the covariance matrix of the local parameters, θ\theta. In my original proposition, ϕ\phi contains the average variance and the average correlation.

Covariance matrix can be deterministic from \phi, and most often is.

@charlesm93

Would it make sense to use a scientific example for the case study?

Better to have first working simulation example with speedup, and where you can show the performance with respect to increasing correlation. You could make a plot like in N_eff BDA3 vs. Stan - #21 by avehtari , just use that same multivariate normal as latent function, e.g., for log-intensity of Poisson.

You can generate Poisson distributed data from your model, and you can check how the speed changes when you vary the parameters of data generating model.