Bindoff, A. D.
2021-04-15
Penalized regression methods are used to select variables for prediction models (or diagnostic models). In some fields the data may be very high-dimensional, resulting in very large model matrixes. Frequently, users must fit many boostrap replications to obtain statistics such as confidence intervals. This can quickly exceed memory and computation constraints, even using sparse matrix representations.
The boot.glmnet R package aims to extend the popular glmnet R
package [1] by providing functions to fit bootstrap replications in
blocks which can be computed sequentially to be saved and combined, or
computed in parallel using the mc_boot_glmnet function which takes
advantage of the future.apply R package [2] to distribute computing
in all supported environments (including Windows).
Additionally, boot.glmnet will calculate the proportion of replicates
for which a predictor term is selected via LASSO (or elastic-net), and
return an approximate confidence interval around this.
Planned extensions include cluster-robust bootstrap methods including cluster wild-bootstrap.
Methods are provided for conveniently summarising results above significance thresholds and plotting results.
TBA
[1] Jerome Friedman, Trevor Hastie, Robert Tibshirani (2010). Regularization Paths for Generalized Linear Models via Coordinate Descent. Journal of Statistical Software, 33(1), 1-22. URL https://www.jstatsoft.org/v33/i01/.
[2] H. Bengtsson, A Unifying Framework for Parallel and Distributed Processing in R using Futures, arXiv:2008.00553, 2020