NEWS (original) (raw)

• Updated functions and code examples to work with glmnet >= 4.1-9 (#20).
  Specifically, compare_by_validate() andcompare_by_calibrate() now gain a new argumentrule for selecting lambda in glmnet models and defaults tolambda.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 oflambda.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.

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.