Initial packaging of the BayGMST model. Previously this repository was a
collection of top-level analysis scripts (R_scripts/, utils/) driven by
config.yml; this release converts that logic into a documented, installable
R package with an explicit function API. The original scripts are preserved
under inst/legacy-scripts/ for provenance, changed only where CRAN policy
requires (they no longer clear the workspace, and they restore par()).
transform_forcings() -- ported from the normalization block in
BayGMST_v1.0.R.reduce_proxies() -- ported from utils/PAGES2k_reducedProxy_UNSC.R.
Supports "PCR", "LASSO", "sPCR", and "SPLS". "SIR" is included but
marked experimental (see "Known limitations" below). It applies the
max_na_frac missingness screen before zero-filling gaps, the order of
Barboza et al.'s original script (PAGES2k_reducedProxy_UNSC_OLD.R). A
later edit of that script (PAGES2k_reducedProxy_UNSC.R) zero-filled
first, which hid internal gaps from the screen and let sparse records enter
the regression mostly as zeros; on large proxy pools that reduced proxy
stopped fit_baygmst() from converging. The output's proxies element
lists the records each segment used.fit_baygmst() -- ported from the data-prep and cmdstanr sampling block
in BayGMST_v1.0.R. Returns a baygmst_fit object instead of relying on
variables left in the global environment.reconstruct() -- extracts a tidy reconstruction data.frame from a
baygmst_fit (equivalent to the gmst_reconstruction_data.csv produced by
the original script), without writing to disk.cv_baygmst() -- ported from the k-fold loop in cv_v0.1.R.plot_reconstruction(), plot_trace(), plot_posterior_densities(),
plot_cv() -- return ggplot objects instead of calling ggsave() as a
side effect.print.baygmst_fit(), summary.baygmst_fit(), print.baygmst_cv() S3
methods.plot.baygmst_fit(): plot(fit) reproduces the combined figure of the
original script and the PReSto manuscript (reconstruction on top; grouped
posterior densities of alpha1, the forcing sensitivities, and the AR(1)
coefficients below). Requires 'patchwork' (Suggests).as_baygmst_proxy() -- normalizes the shapes a reduced/composited proxy
series can arrive in (the output of reduce_proxies(), a data.frame, a
bare vector, a time-by-ensemble matrix, or a composite object from
another package such as a 'compositeR' paleoComposite, with ages in yr
BP converted to calendar years and ensembles collapsed to a median
series) into a year-aware baygmst_proxy object. fit_baygmst() and
cv_baygmst() accept these objects (or anything coercible) directly and
align them to years by calendar year; the original bare-vector,
positional interface is unchanged. Ensemble inputs are collapsed before
fitting in this release (compositing uncertainty is not yet propagated);
the object records n_ensemble for a future ensemble-aware pathway.reduce_proxies() output now carries class baygmst_rp (same list
structure as before) with a print() method.summary() on a fitted model now reports convergence diagnostics
(max_rhat, min_ess_bulk) and its print method warns prominently when
max_rhat > 1.05 (surfaced by a Holocene-scale stress test in which a
poorly identified fit previously printed without complaint).plot_trace() and plot_posterior_densities() no longer emit
"Dropping 'draws_df' class" warnings.baygmst-model describing the statistical model (data and
process levels, priors, forcing provenance, and the mapping to the Stan
code), ported from the hand-typeset draft reference manual's
"Statistical Model" section. It explicitly credits the reduced-proxy
Bayesian framework of Barboza et al. (2014, 2019) and Wang (2020).load_proxies()'s match.arg() choices omitted "sPCR", even though
config.yml and the bundled data/barboza_rps/RP_new_All_sPCR.csv both
support it -- selecting sPCR via the original script would error.
reduce_proxies(method = "sPCR") now works./Users/tylerbagwell/Documents/GitHub/BayGMST_R/data/forcings_with_prediction_HanWang.csv
from the cross-validation logic. That block (Forcings.projections/df_prj)
was also dead code in the original script -- built but never referenced
again -- so it was not ported.rm(list = ls()) call at the top of
utils/PAGES2k_reducedProxy_UNSC.R. A package function must never clear its
caller's global environment.The following were reviewed and signed off by the package authors before the first CRAN release:
y[NT_mis]/z[NT_mis] indexing, kept as is. In
src/stan/baygmst.stan, the observed-period fitted-value block uses the
count NT_mis as a vector index at t == 1. Because missing
(pre-instrumental) years always occupy positions 1:NT_mis, this is
numerically equivalent to y[idx_mis[NT_mis]]/z[idx_mis[NT_mis]] and is
not a live bug. It would be wrong only if missing and observed years were
interleaved, and it affects only y_ins_fitted[1], a posterior-predictive
diagnostic, not the parameter posteriors or the reconstruction y_mis.vignette("baygmst-model").r2_cv, rather than the plug-in estimate of the original cv_v0.1.R.
The two are close but not identical.reduce_proxies(method = "SIR") is experimental and unvalidated and
warns on use. The original SIR implementation depended on legacy
libraries that are no longer available, so it could not be used as a
reference.vol_coef = 25 (W m^-2 per unit AOD) is
attributed to Hansen et al. (2005); the value is corroborated via IPCC AR5.library()-loaded by the original scripts (car, fda,
ggmap, maps) had no evident corresponding usage in the code as written,
so they were not carried into Imports/Suggests.data/HadCRUT.5.1.0.0.analysis.anomalies.ensemble_mean.nc (31 MB) was not
read by any ported function and has been removed from the repository tree
(it remains in git history), along with the compiled CmdStan binaries in
inst/legacy-scripts/ and the outputs/ build artifacts. The legacy
config.yml moved to inst/legacy-scripts/config.yml; no package
function reads it.reference-manual/) was retired.
Its model description lives in vignette("baygmst-model"), and the
reference manual is the one R builds from man/.