General best practices

Building models#

• If possible, build models using JAX
• Make sure code is modular and well-documented
  • Components that could be reused should be made into functions
• Contribute model code to behavioural_modelling if appropriate

Fitting models#

• See this paper for some useful advice on fitting models
• Model parameters should be estimated using a hierarchical Bayesian approach, with posterior samples obtained using MCMC
  • MCMC should be run using Numpyro (unless this isn’t possible for any reason)
  • Posteriors should be checked using traceplots and other diagnostics (e.g., Rhat)
• We are in the process of evaluating simulation-based inference (SBI) as an alternative model-fitting procedure
• Models should be properly evaluated using parameter and model recovery checks