The hardware and bandwidth for this mirror is donated by dogado GmbH, the Webhosting and Full Service-Cloud Provider. Check out our Wordpress Tutorial.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]dogado.de.
Improved CRAN compatibility. Updated package metadata and documentation following CRAN review comments, including expanding Markov chain Monte Carlo
DESCRIPTION file.Modernized examples. Reworked documentation
examples so lightweight examples are executable and self-contained.
Replaced unnecessary \dontrun{} blocks with executable
examples or \donttest{} where appropriate.
Protected computationally intensive examples.
Kept examples that require fitting Bayesian hierarchical models with
Stan inside \dontrun{} because they are not suitable for
routine CRAN example checks.
Improved method documentation examples. Added
small example objects for selected shrinkr_fit and
mixture-method documentation so examples no longer depend on objects
created in other help pages.
federated_learning vignette to save and restore graphical
settings after calls to par(), avoiding persistent changes
to the user’s plotting environment.vignette("map_prior_with_beastt"). Demonstrates
how to build a robust meta-analytic-predictive (MAP) prior by pairing
shrinkr with the beastt package. shrinkr runs the
hierarchical meta-analysis across historical control arms and returns
the MAP as a distributional object, which
beastt then robustifies (robustify_norm()) and
combines with the internal control arm (calc_post_norm())
to form the control posterior and report an effective sample size. The
worked example uses a continuous outcome with known SD, and the prior is
calibrated with sample_prior_predictive() /
prior_pairwise_differences().R/utils.R. The branch handling
dist_inverse_gamma priors on tau was reading
from params$shape and params$scale, but
distributional::parameters() returns those fields as
s and r. The check
!is.null(params$shape) && !is.null(params$scale)
was always FALSE, so the branch silently failed and downstream code
relied on a format-string regex fallback to rescue the parsing.
End-to-end posteriors were correct (the fallback happened to parse
format() output identically to what the explicit branch
should have produced), but the code path was fragile. The branch now
reads params$s / params$r from
parameters(tau_raw) (the unwrapped distribution, not the
possibly-truncated wrapper) and dispatches correctly without needing the
fallback.R/utils.R. The dist_uniform branch
was reading from tau_raw$min and tau_raw$max,
but distributional stores the bounds as l and
u. Same pattern as the IG bug: branch silently failed,
format-string fallback rescued it. The branch now reads
tau_raw$l / tau_raw$u and dispatches
correctly.stan_code() generic and
methods from R/prior_system.R. The generic and its
9 distributional-family methods were vestiges of an earlier, abandoned
design where Stan code was assembled at runtime from prior strings.
shrinkr’s actual workflow handles all prior dispatch internally inside
the precompiled Stan model
(inst/stan/stage2_shrinkage.stan) via integer prior codes
and parameter arrays. The dead code path was never reached at runtime;
removing it drops ~120 lines of source, 9 S3method entries
from NAMESPACE, and 10 .Rd files from
man/. prior_mixture() and
prior_spike_slab() remain unchanged.R/utils.R. Roughly 85 lines across the
mu and tau translation paths used
format(prior) string parsing as a last-resort dispatch when
the explicit class-based branches missed. With the inverse-gamma fix
above, every supported prior now dispatches via class-based checks,
making the fallback redundant. Removing it leaves a cleaner, more
transparent failure mode: an unsupported prior raises a clear error
rather than silently passing through a regex parser.LICENSE field in DESCRIPTION now reads
GPL (>= 3) and the full license text is shipped in
LICENSE. This formalizes the terms under which the package
may be used, modified, and redistributed.compute_theta_contrasts() to
theta_contrasts() for consistency with the other
theta-family functions (extract_theta(),
summarise_theta()). The old name has been removed; update
calls accordingly.summarise_mu_tau() /
summarize_mu_tau(): Summarize hyperparameters
(mu, tau, tau_squared) in the
same format as summarise_theta(). Returns posterior mean,
sd, quantiles, and convergence diagnostics (rhat,
ess_bulk, ess_tail) when multiple chains are
available. Closes a naming-symmetry gap: previously
extract_mu_tau() existed but had no summary
counterpart.array[N] type name form in
inst/stan/stage2_shrinkage.stan. The function signature for
builtin_tau_prior_lpdf() and the tau_params /
custom_params data declarations previously used the
pre-2.26 syntax (real[] params,
real tau_params[6], real custom_params[10]),
which emitted stanc3 deprecation warnings and will become errors in a
future Stan release. No user-visible change in behavior; R-side data
passing is unaffected.std_normal() prior on the inactive
parameterization vector (z when centered,
theta_c when non-centered). In 0.4.0, the inactive vector
had no sampling statement, giving it an implicit improper uniform prior.
This caused Stan’s NUTS sampler to waste effort exploring unbounded
space, significantly slowing sampling. The fix constrains the inactive
vector without affecting the model.Removed parse_priors(): The legacy
formula-based prior interface has been removed. Use
distributional objects (e.g., dist_normal(),
dist_student_t()) directly with shrink() via
the hierarchical_priors argument.
Removed shrinkage_factor from Stan
generated quantities. This quantity (tau^2 / (tau^2 + 1))
was misleading because it assumed standardized Stage 1 variance. Use
tau_squared directly for heterogeneity assessment.
Removed group column from
fit_mixture() output. The column was always 1L
and served no purpose. The variable column identifies
subgroups.
prior_mixture() and
prior_spike_slab() now return native
distributional objects via
distributional::dist_mixture(). This replaces the previous
custom dist_mixture class. All distributional
operations (sampling, density, format, quantiles) work automatically.
Code that accessed $components or $weights
fields directly will need updating.
prior_spike_slab() must be truncated for
tau: Since tau is a scale parameter, spike-and-slab priors must
be wrapped with dist_truncated(..., lower = 0).
sample_prior_predictive() now errors (rather than warns) if
tau draws are negative. Example:
tau = dist_truncated(prior_spike_slab(), lower = 0).
Mixture prior support now works end-to-end:
prior_mixture() and prior_spike_slab() now
correctly flow through to Stan for both mu and
tau priors. Previously, these were exported but not wired
into the .coerce_priors_to_stan() pipeline.
Truncated mu priors: mu now
supports truncation via
dist_truncated(dist_normal(0, 5), lower = 0). This enables
constraining the global mean to a scientifically meaningful
range.
prior_pairwise_differences(): New
function to compute the prior-implied distribution of |theta_i -
theta_j| from prior predictive samples. Includes print()
and plot() S3 methods. The pooled view uses a skyblue
density (matching the hyperparameter panel) and the by-pair view uses
lightcoral violins (matching the theta panel). Useful for calibrating
priors following the recommendation to inspect pairwise difference
distributions before fitting.
Re-exported as_draws_df() from the
posterior package so users can call
as_draws_df(fit) without explicitly loading
posterior.
Safe Cholesky factorization:
.as_chol_factor() now uses a jitter-and-retry strategy
(1e-10 through 1e-4) with Matrix::nearPD() fallback instead
of passing a non-lower-triangular eigen decomposition to Stan. The old
fallback could silently produce invalid cholesky_factor_cov
input.
Stan model no longer exposes unused parameters:
The inactive parameterization (z when centered,
theta_c when non-centered) is constrained with a
std_normal() prior so it doesn’t appear as a meaningful
parameter in output.
Fixed all vignette references: Standardized to
underscore naming (getting_started,
brms_integration, tidy_bayesian_workflow,
federated_learning) throughout R docs, man pages,
vignettes, and README. Removed references to nonexistent vignettes
(mathematical-foundation, two_stage_demo,
tidybayes-integration).
Mixture component extraction handles truncated
normals: The internal .extract_normal_params()
helper correctly unwraps dist_truncated(dist_normal()) to
extract the underlying mean and sd for Stan mixture priors. This is
needed for dist_truncated(prior_spike_slab(), lower = 0) to
work.
test-prior-system.R with 41 tests
covering:
prior_mixture() and prior_spike_slab()
object creation and sampling.coerce_priors_to_stan() for all mu types (Normal,
Student-t, truncated, mixture).coerce_priors_to_stan() for all tau types (8 built-in
+ mixture + error cases).as_chol_factor() safety (positive-definite,
near-singular, bad matrices)sample_prior_predictive() validation and tau positivity
enforcementprior_pairwise_differences() structure, edge cases, and
probability calibrationfit_mixture() output structure (no group
column, quantiles present)shrink() with spike-and-slab tau, truncated
mu, mixture mushrinkage_factor is absent from
outputtest-shrinkr.R: Updated mixture prior tests for
distributional::dist_mixturetest-extraction.R: Updated
as.data.frame.shrinkr_mixture to expect no
group columnREADME substantially improved:
K_max controls and how mclust selects K
via BICimplied_range in prior predictive
sectionfit$fitGetting started vignette: Added section on
prior_pairwise_differences() with explanation of clinical
calibration for pairwise |theta_i - theta_j|.
Package version references synchronized to 0.4.0 throughout.
vignette("federated_learning")). The vignette
covers:
plot.shrinkr_fit() default visualization:
This release focuses on core functionality with a clean, stable API for internal use at GSK.
distributional package for standard prior families:
fit_mixture(): Approximate Stage 1 posteriors with
Gaussian mixture modelsshrink(): Apply hierarchical shrinkage with custom
priorsprior_spike_slab(): Create spike-and-slab mixture
priors for testing homogeneityprior_mixture(): Create custom mixture priors with
multiple componentssample_prior_predictive(): Generate prior predictive
samples for prior elicitationdistributional package syntaxsample_prior_predictive()extract_mu_tau(): Extract global parameter estimates
(mean and heterogeneity)summarise_theta() / summarize_theta():
Summarize group-level estimates with credible intervalscompute_theta_contrasts(): Compute pairwise group
comparisonsplot() methods: Visualize shrinkage effects and mixture
approximation qualityThese binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.
Health stats visible at Monitor.