smooth_model = "random_walk" made the default in
survextrap(), following Timmins et
al..rmst() and related functions are now faster, with the
new Gauss-Legendre integration method. Thanks to Iain Timmins.disc_rate to
mean and rmst() functions (PR #12)irmst, hazard_ratio and
hrtime tidied. Fix of bug in irmst introduced
in v0.8.9, and all now have a sample argument.standardise_to() function.loo no longer ignores covariates on the cure
probability in cure models.
More internal refactoring of code to produce outputs.
Cleaned and simplified internal code for model outputs
(rmst, survival etc.).
rmst now outputs a tibble, and credible limits in
the default output are renamed to lower and
upper for consistency with survival and
hazard.
Arrays now allowed for coef argument in
dsurvspline and related functions.
Background hazards in additive hazards models can now be
stratified, using the new backhaz_strata argument to
survextrap(). This argument names stratifying variables
(e.g. age group) that should be in both the background hazards and the
individual and/or external data.
Bug fix for hsurvmspline_wane and
dsurvmspline_wane where offsets were not implemented
correctly. This affected hazard predictions for models that included
both background hazards and treatment effect waning.
smooth_model = "random_walk".Non-proportional hazards models can now be applied to a subset of
the covariates, by supplying a formula as the nonprop
argument to survextrap.
Prior simulation functions such as prior_sample now
require an explicit covariate model, specified through
formula and newdata, rather than a design
matrix X. These functions now fully support nonproportional
hazards models and cure probabilities with associated regression
models.
mspline specification objects
tidied, with new mspline_init and
mspline_list_init functions.rstan version 2.26.mspline_spec bug introduced in 0.8.2.The GitHub
commit on 6 May 2023 implemented several changes that are listed
here. survextrap is now in “beta” status. All major
features are now implemented, but there may still be some lack of
polish.
A new vignette cetuximab, giving an in-depth case
study of a realistic use of survextrap. To be included in a
forthcoming paper.
Improvements to usability and stability of M-splines:
iknots and bknots are replaced by a
single knots, and the lower boundary knot is always fixed
at zero.
add_knots can be used to add user-defined knots to
the default ones. See the case study.
New function mspline_spec to define M-spline knots
based on data, in advance of fitting any models.
M-splines can now be smoother at the final knot, through an
option bsmooth, which is on by default. Thanks to Iain
Timmins for the suggestion.
Incremental restricted mean survival time (irmst)
added.
Any summary of the posterior distribution can be produced by
supplying a list of summary functions (e.g. mean,
quantiles) to output functions (such as rmst
and survival).
Leave-one-out cross validation for external data.
Bug fix for hsurvmspline() (hence
hazard()) which was using the wrong tail end of
psurvmspline().
Waning models no longer ignore non-proportional hazards and cure. Waning vignette merged into the methods vignette.
Leave-one-out cross validation now supports models with background hazards.
More consistent naming of arguments and variables
(e.g. hscale for hazard scale, and hsd for
hazard smoothing parameter).
More information given when printing fitted model objects.
Thorough documentation polishing and code tidying.
See commit messages on GitHub for a history of developments to
survextrap.