coevolve 1.2.1

coevolve 1.2.1

New Features

• Added further documentation about repeated observations models (#120)
• Added vignette "Algorithmic scaling" (#121)
• Models with Gaussian variables and no repeated observations now use a
  conditionally non-centred parameterisation for latent terminal drift, i.e.
  the terminal drift of non-Gaussian variables and of missing Gaussian values.
  Latent drift is written as a standard normal innovation conditional on the
  observed Gaussian residuals of the same taxon. This is a change of variables
  that leaves the model, the posterior, and the pointwise log-likelihood
  unchanged, but removes a funnel between the drift parameters and the latent
  terminal drift that caused poor mixing in mixed Gaussian and non-Gaussian
  models. Thanks to @yuhangxoox for the diagnosis (#124)

Bug Fixes

• Fixed the continuous time intercept term in the expected change along a
  branch, A^-1 (exp(A t) - I) b. Since version 1.0.0 this matrix was
  symmetrised before being multiplied by b, which is only correct when the
  selection matrix A is symmetric (e.g., no cross-selection effects). Both
  the Stan and JAX backends now use the exact expression. In our comparisons
  the selection (A) and drift (Q) parameters were essentially unchanged,
  because b is weakly identified and absorbs most of the difference.
  Posterior estimates of b itself can shift; in one simulated example they
  moved by up to about 0.6 posterior standard deviations
• Fixed coev_plot_predictive_check() when nuts_sampler = "nutpie". The
  JAX backend now returns yrep posterior predictions alongside the other
  model parameters, matching the Stan backend (#118)
• log_lik = TRUE is now supported with nuts_sampler = "nutpie". The JAX
  backend previously ignored the argument and returned no log_lik, so
  loo and waic could not be computed from nutpie fits (#118)
• Fixed prior_only = TRUE for models with Gaussian variables and no
  repeated observations. terminal_drift only received a prior inside the
  likelihood block, leaving it improper when the likelihood was skipped.
  Sampling failed to initialise with nuts_sampler = "nutpie" and was
  unreliable with nuts_sampler = "stan" (#118)
• Fixed the pointwise log_lik and yrep for Gaussian variables with
  missing data. The generated quantities block used the -9999 missing
  value placeholder when constructing residuals, which corrupted the
  conditional densities of the other variables of the same observation