modelInitialization: Information on initial values in a NIMBLE model

In nimble: MCMC, Particle Filtering, and Programmable Hierarchical Modeling

Information on initial values in a NIMBLE model

Description

Having uninitialized nodes in a NIMBLE model can potentially cause some algorithms to fail and can lead to poor performance in others. Here are some general guidelines on how non-intitialized variables can affect performance:

• MCMC will auto-initialize but will do so from the prior distribution. This can cause slow convergence, especially in the case of diffuse priors.

• Likewise, particle filtering methods will initialize top-level parameters from their prior distributions, which can lead to errors or poor performance in these methods.

nimble documentation built on March 18, 2022, 8:03 p.m.