NEWS.md
As of February 2024: * Updated versions of hal9001 and origami in DESCRIPTION to match the latest CRAN releases, resolving bugs related to Matrix v1.6-2 as reported at https://github.com/tlverse/hal9001/issues/109.
As of May 2021: * Updated nonparametric IPW estimation code to match the methods described in https://arxiv.org/abs/2205.05777. * Updated tests, references, and pkgdown site in preparation for JOSS paper.
As of October 2021: * Removed LazyData field in DESCRIPTION since no data directory present. * Removed reference to glmnet in documentation to avoid adding the package to dependencies.
As of October 2021: * Reduced time-intensive nature of unit tests as per CRAN policies. * Added smoothness_orders as a named argument to haldensify, fit_haldensify, and cv_haldensify, with a default of zero. This was previously passed to hal9001::fit_hal via ... arguments. * Submission accepted by and posted to CRAN.
As of September 2021: * Refinements of internal calls to hal9001::fit_hal() in keeping with updates to that package, for compatibility with its v0.4.0 CRAN release. * The smoothness_orders argument of hal9001::fit_hal() previously was set through the ... argument of haldensify; however, it has now been made a named argument to both haldensify and the internal cv_haldensify and fit_haldensify functions. The default is set to zero, for indicator basis functions, which differs from the default of hal9001::fit_hal() as of its v0.4.0 release.
As of April 2021: * Changes to internal calls of hal9001::fit_hal() in order to correctly use the pared-down interface introduced in v0.4.0, contributed by @rachaelvp. * The default for the grid of bins used for discretization of the variable A has been altered to be multiples of sqrt(length(A)).
As of April 2021: * Updates to haldensify arguments (removal of hal_max_degree as a named argument) to simplify and better match use of fit_hal in hal9001 v0.3.0+. This overhaul also included the addition of ... arguments, now passed through haldensify and fit_haldensify to cv_haldensify, allowing all internal calls to hal9001::fit_hal() to specify the same arguments be passed for the fitting of HAL models. * Changes to the default values of the argument n_bins, now setting this to (much) larger values that are themselves based on the sample size. This is in accordance with evidence from simulation experiments indicating that higher values of n_bins lead to significantly improved density estimates. * Addition of argument trim and trim_dens to predict.haldensify to support the use of truncation more transparently. While the default was to set predictions for values of new_A outside the training support to zero, this has been changed to avoid trimming and, when the choice is made to trim the predictions, to set this value to 1/sqrt(length(new_A)). * Addition of a new method print.haldensify for a more user-friendly display of the prediction procedure’s output, including the selected number of bins, the CV-selected choice of the regularization parameter, and the summary of the fitted HAL model.
As of February 2021: * Addition of a method plot.haldensify to simplify visualizing the empirical risks of the sequence of HAL-based conditional density estimators across the grid of the regularization parameter, and necessary changes to the vignette. * Preparation to add an option to visualize the conditional density estimates (of the estimator selected by cross-validation) via a type argument in the plot.haldensify method. Not yet implemented. * Simplification of unit tests to remove unnecessary reliance on dplyr. * Limit re-fitting of HAL model (after CV-selection of tuning parameters) in haldensify() to full-data fit by explicitly passing n_folds = 1. * Avoid cross-validation procedure conditionally when the arguments n_bin and grid_type are fixed; add related assertion check in predict() when haldensify() skips cross-validation (since lambda selection skipped). * Change how long-format repeated measures data is passed around in both haldensify() and predict() to clarify variable passing. * Correct predict() method to truncate small conditional density estimates to a minimum value of [1 / sqrt(n)], based on the prediction set sample size.
As of January 2021: * Addition of argument hal_basis_list to haldensify(), allowing for a HAL basis produced by fit_hal() to be passed into the HAL regression used for density estimation. This facilitates reduced computational overhead when requiring external cross-validation of nuisance functions (e.g., CV-TMLE) as well as working with bootstrap samples. * Addition of argument hal_max_degree to haldensify(), allowing for control of the highest degree of interactions considered in the HAL model for density estimation. Like the above, this can reduce computational overhead. * Fix a minor bug in haldensify() by passing cv_folds to the n_folds argument of fit_hal() when fitting HAL regression for density estimation. Previously, cv_folds was only used in constructing cross-validation (CV) folds for choosing tuning parameters, but the subsequent HAL regression was fiex to use the default number of folds specified in fit_hal() to choose the regularization parameter of the HAL regression for density estimation. Now, both CV to choose density estimation tuning parameters and CV to choose the lasso tuning parameter use the same number of folds. * Addition of argument ... to haldensify() so that arbitrary arguments can be passed to fit_hal() for density estimation, when not already specified as other arguments of the haldensify() constructor. * Remove the unnecessary argument use_future, specifying parallel evaluation in a note instead. * Add an option "all" to the lambda_select argument of the predict() method, allowing for predictions on the full (non-truncated) sequence of lambdas fitted on to be returned. * Change truncation option in predict() method to 1/n instead of zero.
As of January 2021: * Adds support to facilitate convenient marginal density estimation by creating automatically a constant vector when W = NULL is set in haldensify(). * The hal9001 dependency has been upgraded to v0.2.8 of that package, which introduced breaking changes in the names of slots in fitted model objects. * The sequence of HAL models re-fit after identification of the regularization parameter selected by cross-validation has been padded with more aggressive choices of the parameter to ameliorate convergence issues in model fitting. * Re-fitting of the HAL model with cross-validated choices of the number of bins, binning procedure, and regularization sequence has been altered to reuse the regularization sequence provided as input rather than subsetting the sequence to start with the cross-validation selector’s choice of the parameter. Though convenient for undersmoothing haldensify estimates, this subsetting proved problematic for convergence of glmnet(). * The predict() method’s cv_select argument has been replaced in order to better facilitate undersmoothing. The new argument lambda_select defaults to the cross-validation selector but now easily allows access to the sequence of undersmoothed density estimates (less restrictive regularization values). * The names of three slots in the haldensify S3 output class have been changed * grid_type_tune_opt is now grid_type_cvselect, * n_bins_tune_opt is now n_bins_cvselect, and * cv_hal_fits_tune_opt is now cv_tuning_results.
As of December 2020: * Use of plan(transparent) has been changed to plan(sequential) based on ongoing development in the future package ecosystem.
As of June 2020: * A short software paper for inclusion in JOSS has been added.
As of May 2020: * The core cross-validation routine in haldensify for fitting HAL models has been slightly abstracted and moved to the new function fit_haldensify. * The haldensify() wrapper function serves to cross-validate over choices of the histogram binning strategy and the number of bins. * The defaults of haldensify() have been changed based on results of simulation experiments. * The unnecessary argument seed_int in haldensify() has been removed. * Fixes a bug introduced by returning predicted hazards as a vector instead of a matrix. * An argument cv_select, defaulting to TRUE, has been added to the predict method, to make undersmoothing more accessible. * A simple vignette has been added.