NEWS
coevolve 1.3.0
New Features
- Added
scale_trees argument to coev_fit(). This argument allows users to
toggle tree scaling on or off. By default, phylogenetic trees are scaled by the
maximum root age across trees, preserving time scales across samples of trees.
If turned off, phylogenetic trees are not scaled. For backwards compatibility,
models with individual trees should remain unchanged, but models with samples of
trees may return different results due to different data preparation (#127)
Other Changes
- Completed rOpenSci package review and transferred repository to rOpenSci
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
coevolve 1.2.0
New Features
- Added
coev_ancestral_states() for extracting posterior estimates of
ancestral trait values at internal phylogenetic nodes, on either the
latent or response scale (#86)
- Added vignette "Ancestral state reconstruction"
Bug Fixes
- Fixed issue with plotting functions when
nuts_sampler = "nutpie" (#114)
Other Changes
- Updated documentation (#111)
- Fixed cmdstanr location in documentation (#112)
- Responded to rOpenSci package review (#113)
- Added package citation (#116)
coevolve 1.1.0
New Features
- Allowed for single traits (#107)
- Added pure JAX/NumPyro backend via
nuts_sampler = "nutpie".
Uses nutpie's Rust NUTS sampler with JAX gradients for ~5x
faster sampling than Stan on typical models. Requires
pip install jax numpyro nutpie (#109)
coevolve 1.0.0
New Features
- Added
nutpie as an alternative sampler for the Stan models (#92)
- Cached matrix computations when branch lengths are identical (#93)
- Implemented Hilbert-space approximate Gaussian processes for spatial control,
adding the
lon_lat argument and deprecating the dist_mat argument in
coev_fit() (#103)
Bug Fixes
- Fixed issue with
summary() when estimate_residual = FALSE (#95)
Other Changes
- Updated license
- Submitted to rOpenSci for package review
- Linted package using lintr
- Implemented srr compliance checks
- Added continuous integration checks using GitHub Actions
- Reduced cyclomatic complexity for some functions
- Re-factored Stan code generation to use
whisker templates (#99)
- Implemented automatic test fixture re-generation in GitHub Actions (#100)
coevolve 0.1.0