Bayesian Additive Regression Trees (BART) provide flexible nonparametric modeling of covariates for continuous, binary, categorical and time-to-event outcomes. For more information see Sparapani, Spanbauer and McCulloch
Changes and New Features in 2.4 (2019-04-10): * change: per CRAN policy, dynamic libraries are no longer "stripped" on Linux
Changes and New Features in 2.3 (2019-03-27): * new feature: adding arguments to surv.pre.bart, surv.bart and mc.surv.bart to fine-tune grid of time points and automate creation of time dependent covariates. These are convenience features to make multi-state models easier to handle; see new demo leuk
* change: arguments to rtnorm and rtgamma more user friendly
* new feature: re-organized vignettes into a single vignette
* new feature: gbart now calculates log pseudo-marginal
likelihood (LPML) for computing pseudo-Bayes factors
* new feature: new Generalized BART Mixed Models, see the
function gbmm
Changes and New Features in 2.2 (2019-01-22):
* bug fix: fix typo in size of theta grid for sparse prior
* new feature: Multinomial BART, mbart2, (suitable for
cases with more categories) based on the original mbart
implementation but inspired by the logit transformation;
nevertheless, both logit and probit are available and,
of course, probit is much faster
Changes and New Features in 2.1 (2018-11-28):
* to meet current CRAN guidelines, replaced CXX1X and CXX1XSTD
configure/autoconf macros with CXX11 and CXX11STD respectively
Changes and New Features in 2.0 (2018-11-12):
* new feature: Multinomial BART, mbart, (suitable for cases
with relatively fewer categories) replaced with a new
conditional probability implementation which allows the user
to choose probit or logit BART; of course, probit BART is
much faster
* new feature: if lambda is specified as 0, then sigma is
considered to be fixed and known at the value sigest
and, therefore, not sampled
* bug fix: fixed single column x.test bug
Changes and New Features in 1.9 (2018-08-17):
* bug fix: off by one error fixed in robust Gamma
generator for sparse Dirichlet prior
* new feature: abart/mc.abart computes a variant
of the Accelerated Failue Time model based on BART
* new feature: for x.train/x.test with missing data elements,
gbart will singly impute them with hot decking.
Since mc.gbart runs multiple gbart threads in parallel,
mc.gbart performs multiple imputation with hot decking,
i.e., a separate imputation for each thread.
Changes and New Features in 1.8 (2018-06-30):
* bug fix: fix typo in the recur.pwbart() which
prevented predict() from working when OpenMP
was not available
Changes and New Features in 1.7 (2018-06-08):
* enhancement: generalized, or generic, BART: gbart/mc.gbart
unites continuous and binary BART in one function call
re-based time-to-event BARTs on gbart as well
* enhancement: binaryOffset=NULL specifies
binaryOffset=qXXXX(mean(y.train)) for pbart/mc.pbart,
lbart/mc.lbart, mbart/mc.mbart; offset=NULL does the
same for gbart/mc.gbart, surv.bart/mc.surv.bart,
recur.bart/mc.recur.bart, crisk.bart/mc.crisk.bart
and crisk2.bart/mc.crisk2.bart (note: competing
cause 2 is handled analogously for offset2=NULL)
* enhancement: multinomial BART rebased on probit BART for
computational efficiency
* bug fix: several corrections in probit and logit BART.
Note that this may change your results for binary and
time-to-event outcomes. For probit BART, the correction
generally leads to a small change in the results. However,
the logit BART correction may lead to more substantial
changes.
* doc fix: correct docs for the binary case in pbart/mc.pbart,
lbart and mbart; and correct docs for the numeric case in
wbart/mc.wbart
* enhancement: robust Gamma generation for small scale parameter
* enhancement: more robust sparse Dirichlet prior implementation
Changes and New Features in 1.6 (2018-03-19):
* for binary outcomes, new default for ntree=50
(change inadvertently omitted from v1.4 below)
* enhancement: recur.pre.bart, recur.bart and mc.recur.bart
can now handle NA entries in the times and delta matrices
* enhancement: for time-to-event outcomes, new optional
K parameter which coarsens time per the quantiles
1/K, 2/K, ..., K/K.
* bug fix: x.test/x.test2 now properly transposed if needed
for post-processing
* bug fix: sparse Dirichlet prior now corrected for
random theta update. Thanks to Antonio Linero for
the detailed bug report.
Changes and New Features in 1.5 (2018-02-08):
* bug fix: ambiguous call of floor surrounding integer division
* bug fix: x.test is not an argument of recur.pre.bart
Changes and New Features in 1.4 (2018-02-02):
* for binary outcomes, new default for ntree=50
* fixed library bloat on Linux with strip
* x.train and x.test can be supplied as data.frames
which contain factors as stated in the documentation
* cutpoints now based on data itself, i.e., binary or
ordinal covariates. Similarly, you can request
quantiles via the usequants setting.
* sparse variable selection now available with the
sparse=TRUE argument; see the documentation
* new vignettes
* new function, mc.lbart, for logit BART in parallel
* mbart updated to equivalent functionality as other functions
* new function, mc.mbart, for Multinomial BART in parallel
Changes and New Features in 1.3 (2017-09-18):
* new examples in demo directory
* return ndpost values rather ndpost/keepevery
* for calling BART directly from C++, you can
now use the RNG provided by Rmath or the STL random class
see the improved example in cxx-ex
* new predict S3 methods, see predict.wbart and other
predict variants
* Added Geweke diagnostics for pbart, surv.bart, etc.
See gewekediag which is adapted from the coda package
* logit BART added for binary outcomes; see lbart
* Multinomial BART added for categorical outcomes; see mbart
Changes and New Features in 1.2 (2017-04-30):
* you can now call BART directly from C++ with the Rmath library
see new header rn.h and the example in cxx-ex
Changes and New Features in 1.1 (2017-04-13):
* No user visible changes: bug-fix release
Changes and New Features in 1.0 (2017-04-07):
* First release on CRAN