Sampler settings

For debugging and developing models, where it is important to get rapid feedback, or for educational use, a low number of warmup and regular samples, e.g. 300 / 500, sampled on a single chain is often good enough. For production / clinical use, however, we would recommend a larger number of samples, and sampled on multiple chains. With more chains, you will also be able to calculate additional diagnostics about convergence.

In our internal testing of several common PK and PK/PD problems, the influence of adapt_delta was limited, a lower value of 0.8 giving very similar precision as 0.95, at a slightly lower computational burden.

If speed of computation is critical for the intended application, it would be appropriate to perform a benchmark study for the specific problem to identify settings that satisfy both the time and precision constraints.

The following default settings for get_mcmc_posterior() override the defaults used by CmdStanR, and are aimed at intermediate precision and faster computation:

• iter_warmup: 500
• iter_sampling: 500
• chains: 1