Bagged ("leafy") partDSA no longer errors when
leafy.random.num.variables.per.split is greater than or equal to the
number of predictors; the random variable subset ("mtry") is now capped
at the number of predictors.
trim() on a partDSA object now actually trims the inherited fields
(IkPn, coefficients, IkPn.risks, var.importance). Previously the
result of the inherited dsa trim method was discarded, so only the
partDSA-specific fields were shortened.
loss.function = "DR" for censored outcomes now fails with a clear
message (it was never implemented and previously raised a cryptic
"object 'get.overall.risk' not found").
Survival fits no longer print a spurious n == length(y)%/%2 is not TRUE
error to the console. The internal per-partition outcome list relied on
length(<Surv>) returning 2n; modern versions of the survival
package return n, which silently disabled $datalist (and therefore
dumpDSA() / showDSA() for survival outcomes). It now uses NROW().
The Integrated Brier Score reported for multiple time points is now a
proper, normalized score. Previously the per-time-point Brier scores
were summed (default IBS.wt = rep(1, length(brier.vec))), inflating
the reported value by roughly the number of time points and pushing it
outside the valid [0, 0.25] range in both mean.cv.risk.DSA and
test.set.risk.DSA. The default now uses trapezoidal weights that sum
to 1 and normalize by the time range. Single-time-point behavior is
unchanged; explicitly supplied IBS.wt is respected (and aligned to
the now-sorted brier.vec).
DESCRIPTION from the deprecated Author:/Maintainer:
fields to Authors@R.inst/CITATION from the old-style citEntry() / personList()
to bibentry() / person().split.fx() to find.split.fx() so it is
no longer mistaken for an unregistered S3 method of the base split()
generic.testthat test suite (tests/testthat/).\Address/\URL so the JSS link/to/webpage placeholder is
no longer emitted into the PDFs. Removed a redundant data(GBSG2) call
in the survival vignette.