Fix documentation NOTEs

Fix CRAN check NOTEs in documentation due to new \link{} handling, and some new NOTEs due to differences between \itemize and \describe.

Add cluster bootstrap

Add a cluster bootstrap for correlated data.

Enhanced VIM point estimation

For inference on VIM values to be valid under the zero-importance null hypothesis, we need this inference to be based on predictiveness (of the two groups of variables that we're comparing) that is estimated on independent splits of the data.

However, for point estimation, this is not required. Until now, the only option has been for the final VIM point estimate to be based on this sample-splitting. Now, you have the option for the final point estimate to be based on the entire dataset (setting final_point_estimate = "full") or for the estimate to be based on averaging the two split-specific VIMs (setting final_point_estimate = "average").

In large datasets (with many rows), there will not be much difference between these three options. However, using more data may help stabilize the final point estimate in smaller sample-size settings. Care must be taken in interpretation, however: inference is still based on sample-splitting, so an identity-link Wald confidence interval will not be centered at the final point estimate if "full" or "average" is used.

vimp 2.3.0

Major changes

• Predictiveness measures now have their own S3 class, which makes internal code cleaner and facilitates simpler addition of new predictiveness measures.
• In this version, the default return value of extract_sampled_split_predictions is a vector, not a list. This facilitates proper use in the new version of the package.

Minor changes

• You can now specify truncate = FALSE in vimp_ci

Specify 'method' and 'family' in outer functions

• Specify method and family for weighted EIF estimation within outer functions (vim, cv_vim, sp_vim) rather than the measure* functions. This allows compatibility for binary outcomes.
• Added a vignette for coarsened-data settings.

Allow parallel computation of CV.SuperLearners

Allow parallel computation in CV.SuperLearner, but not in SuperLearner.

Improved bootstrap and coarsened-data behavior

Allow further specification of the bootstrap (e.g., percentile); update documentation and internal checks for coarsened-data settings.

Allow odd number of folds for CV-VIM with precomputed regressions

v2.2.5

Update the website

Add data

In this release, we:

• added a dataset, vrc01, containing neutralization sensitivity values to the broadly neutralizing antibody VRC01 and HIV viral sequence features
• updated vignettes to use vrc01
• revised how odd numbers of cross-fitting folds are handled

Finalize standard errors

Updated computation of standard errors. This release is harmonized with CRAN release 2.2.3.