Updated functions and code examples to work with glmnet >= 4.1-9 (#20).
Specifically, compare_by_validate() and
compare_by_calibrate() now gain a new argument
rule for selecting lambda in glmnet models and defaults to
lambda.min. Previously, this is not tunable and defaulted
to lambda.1se. Code examples involving glmnet models are
updated to use the lambda.min rule instead of
lambda.1se so they won’t generate null models.
In glmnet 4.1-9, using the lambda.1se rule for choosing
lambda from Cox model cross-validation results can generate null models
more easily. Previously, cross-validation errors for Cox models were
normalized by the sum of weights times event indicators per fold. In
glmnet 4.1-9, cross-validation errors are normalized by the sum of all
weights per fold. This change means that CV errors are divided by larger
values (since censored observations now contribute to normalization),
and it produces smaller CV error values and potentially smaller standard
errors. It could make the 1SE rule more conservative and lead to
selecting larger penalty values, hence sparser models or complete null
models. This particularly affects datasets with low event rates or many
censored observations, where the denominator change is most
significant.
_R_USE_STRICT_R_HEADERS_ (upcoming default in R 4.5.0)
(#13):
STRICT_R_HEADERS.Calloc and Free with
R_Calloc and R_Free.float.h for FLT_EPSILON.ncvsurv() in a pathwise manner instead of with
a single value of lambda.survAUC. Ported the essential
C code for computing time-dependent AUC and fixed the build issues in
r-devel.This version is a major refactor of the package, with several technical adjustments to improve the functional interface, code structure, and execution performance. As a result, a few critical API-breaking changes have been made. Please update your previous code that calls hdnom accordingly. For the detailed changes, please check the updated items below.
hdcox.*() are renamed as fit_*(),
hdnom.nomogram() is renamed as as_nomogram(),
hdnom.validate() is renamed as validate(), and
so on.rms, by reusing a minimal set
of code from rms for nomogram construction and plotting.
This results in clearer code structure, better maintainability, and
faster package installation/loading speed. Also removed other
non-essential package dependencies.print functions are now returned
invisbily, to make it easier to use them in a pipe.fit$model, and the selected “optimal” hyperparameters can
be accessed by fit$lambda. The model type is now stored
explicitly as fit$type.as_nomogram (previously hdnom.nomogram())
now accepts the fitted model objects directly instead of the
$model component. It now will recognize the model type
automatically, thus the previous arguments model.type has
been deprecated. so that it is easier to chain the function calls
together using magrittr.as_nomogram, the previous ddist
argument is not needed anymore and has been removed. There is also no
more need to set a datadist object as a into the global
options variable (which was required in the rms user
flow).theme_hdnom() and applies it to
most of the validation, calibration, and comparison plots for a
consistent, cleaner look across plots within the package.glmnet.survcurve(), ncvreg.survcurve(),
penalized.survcurve()) and Breslow baseline hazard
estimator functions (glmnet.basesurv(),
ncvreg.basesurv(), penalized.basesurv()).hdnom.calibrate().README.md.lambda1 and lambda2
instead of a single “lambda” are now required to fit, validate, and
calibrate fused lasso models.lambda in hdnom.nomogram is
no longer needed and has been deprecated.eps and max.iter
for MCP and SCAD penalty related models. Setting the default values to
be 1e-4 and 10000, which is consistent with
ncvreg 3.8-0.hdnom.kmplot() under ggplot2 2.2.0, which is caused by a
previous workaround for a bug introduced in ggplot2 2.1.0.max.iter for
ncvsurv to a substantially higher value (5e+4).ncvsurv under
ncvreg >= 3.7-0.ylim for
plot.hdnom.validate(),
plot.hdnom.external.validate(), and
plot.hdnom.compare.validate() (#4).hdnom.compare.validate() for model
comparison by validationhdnom.compare.calibrate() for model
comparison by calibrationhdnom.external.validate() for external
validationhdnom.external.calibrate() for external
calibrationpredict and print methods for
hdcox.model objectshdnom.kmplot(): Kaplan-Meier analysis for
risk groups using internal/external calibration resultshdnom.logrank(): Log-rank test for risk
groups using internal/external calibration resultsRhdcox.*() functions. Make examples compatible with ncvreg
3.5-0, which refined CV implementation for survival models and improved
computation speed.Support five more high-dimensional penalized Cox model types:
hdnom.validate(),
hdnom.calibrate(), hdcox.aenet(), and
hdcox.enet() by reducing resampling times.parallel to hdcox.aenet()
and hdcox.enet() to enable or disable the use of parallel
parameter tuning.
Need a high-speed mirror for your open-source project?
Contact our mirror admin team at info@clientvps.com.
This archive is provided as a free public service to the community.
Proudly supported by infrastructure from VPSPulse , RxServers , BuyNumber , UnitVPS , OffshoreName and secure payment technology by ArionPay.