Requires dbarts (>= 1.0-0) and a C++20 compiler. Results differ from earlier versions, including under a fixed seed.
Stan is no longer used. The parametric component is drawn with the
WALNUTS sampler from the Walnutpie library. The priors, QR,
adapt_delta and init_r are now arguments of
stan4bart(), named as in rstanarm; stan_args
is deprecated and routes its entries to them with a warning.
The NUTS tuning arguments
(adapt_gamma, adapt_kappa, adapt_t0,
adapt_init_buffer, adapt_term_buffer,
adapt_window, stepsize, stepsize_jitter,
max_treedepth) are ignored with a warning. The shrinkage
priors hs, hs_plus, lasso, laplace and
product_normal are no longer supported and prior_intercept
is dropped with a warning.
bart_args follows dbarts 1.0-0: node.prior is
now leaf.prior (an error naming the new spelling),
rngSeed is now seed (a warning), and the residual
model arguments (sigma, resid.prior,
resid.dist) are refused. A name in bart_args that
nothing recognizes is now an error; before it was silently ignored.
Warmup draws are no longer stored by default. A per-chain summary
of the sampler tuning is returned as adaptation instead;
use save_warmup = TRUE to keep the draws.
For a binary response, weights that are all 0 or 1 now select
the rows in use and any other weights are an error.
mvbart fits multivariate BART to correlated continuous
outcomes with a shared predictor set.
store = "trees" keeps the trees rather than the full matrix
of BART fits, which shrinks large fit objects considerably;
extract, fitted and predict work either way.
print, summary and residuals methods for
fitted models; the first two warn when the sampler looks
under-warmed.
skip accepts a "stan" element to take several
parametric transitions per BART sweep, which improves mixing of the
random effect standard deviations.
Factors in the bart component are split by level subsets
rather than indicator columns, so extract(type = "varcount")
reports one row per factor.
Compiles and fits under Eigen 5, which a coming release of RcppEigen brings; the package failed to compile there, and once compiled crashed when fitting. Report and fix thanks to Dirk Eddelbuettel.
No longer uses rpath on Window64 arm. PR thanks to
Andrew Johnson.
extract argument now works correctly with varcount. Bug
report thanks to Joshua Bon.
predict now works with single chains. Bug report thanks to
github user Pentaonia (Louis).
An experimental implementation of callbacks has been added as has the option to avoid writing out the full set of Stan and BART samples. This allows a model to be fit just saving samples of the quantity of interested.
treated argument now works correctly with logical vectors.
extract now accepts as a type "trees", which
allows for easier inspection of models fit with
bart_args = list("keepTrees" = TRUE).
Initial public release.