NONMEM Bayes models in bbr

This page provides an overview of the basic structure of NONMEM Bayes models in bbr. The main entry point for interacting with NONMEM Bayes models is the bbi_nmbayes_model object.

Details

Model Structure

NONMEM models in bbr have one control stream. Executing the model produces an output directory. NONMEM Bayes models, on the other hand, have a more complicated model structure. The user still defines only one control stream. Underneath, this top-level control stream is then used to derive the control stream of multiple submodels, one for the initial estimates and one for each chain.

For a control stream with the base name run, submit_model() will create the following submodels:

<run>/init/ - A model that generates the initial estimates.

<run>/<run>-<chain>/ - A model for each chain that generates samples using BAYES or NUTS estimation methods.

Defining a NONMEM Bayes model

The first step to defining a NONMEM Bayes model will often be to call copy_model_as_nmbayes() on an existing non-Bayes NONMEM model. This creates the new model with bbr::copy_model_from() and then sets the model type to "nmbayes". Alternatively you can also create a fresh NONMEM Bayes model with new_model(..., .model_type = "nmbayes"). (As with regular NONMEM models, a control stream must already exist.)

The control stream must define the following two records:

• a METHOD=CHAIN estimation record that defines how to generate the initial estimates for each chain. ISAMPLE should be set to 0. The NITER option controls the number of chains.

• a METHOD=BAYES or METHOD=NUTS estimation record to generate posterior samples.

Here's an example that defines four chains of 500 samples:

$EST METHOD=CHAIN FILE=1100.chn NSAMPLE=4 ISAMPLE=0 SEED=1
     CTYPE=0 IACCEPT=0.3 DF=10 DFS=0

$EST METHOD=NUTS AUTO=2 CTYPE=0 OLKJDF=2 OVARF=1 SEED=1 NBURN=250 NITER=500
     NUTS_DELTA=0.95 PRINT=10 MSFO=./1100.msf RANMETHOD=P PARAFPRINT=10000
     BAYES_PHI_STORE=1