Skip to content

censored 0.4.0

Latest

Choose a tag to compare

@hfrick hfrick released this 13 Nov 17:24

Breaking changes

  • bag_tree() with the "rpart" engine no longer accepts case weights. Fit the model without case weights instead (#390).

  • bag_tree() with the "rpart" engine now passes tree_depth, min_n, and cost_complexity on to the trees, so default fits now use the bag_tree() defaults of cost_complexity = 0 and min_n = 2. To restore the previous fits, set cost_complexity = 0.01 and min_n = 20 (#406).

  • boost_tree() with the "mboost" engine no longer accepts case weights because mboost doesn't predict from a weighted fit. Fit the model without case weights instead (#363).

  • boost_tree() with the "mboost" engine no longer allows setting nu via set_engine(). Use learn_rate instead (#395).

  • boost_tree() with the "mboost" engine now warns and ignores control and tree_controls when set via set_engine(). Previously they silently overrode the main arguments, such as trees and tree_depth. Use the main arguments of boost_tree() instead (#414).

  • survival_prob_*() functions now error when given the defunct time argument. Use eval_time instead (#419).

  • survival_prob_coxph() no longer has the defunct x argument. Use object instead (#419).

New features

  • boost_tree() with the "mboost" engine now passes learn_rate on to mboost::boost_control(nu), making it tunable. Previously it was silently ignored (#395).

  • multi_predict() is now available for boost_tree() with the "mboost" engine over the trees submodel parameter (#290).

  • null_model() gains a "survival" engine for censored regression, fitting a Kaplan-Meier curve via survival::survfit() (#353).

  • rand_forest() gains a "randomForestSRC" engine for censored regression, fitting a survival random forest via randomForestSRC::rfsrc() (#130).

  • rand_forest() gains a "ranger" engine for censored regression, fitting a survival random forest via ranger::ranger() (#131).

Minor improvements and fixes

  • bag_tree() with the "rpart" engine now passes tree_depth, min_n, and cost_complexity, as well as other rpart::rpart.control() arguments set via set_engine(), on to the trees. Previously they were silently ignored (#406).

  • decision_tree() with the "rpart" engine now correctly returns the median survival time of the leaf's Kaplan-Meier curve for type = "time" predictions, instead of rpart's relative event rate (#331).

  • proportional_hazards() with the "glmnet" engine now predicts on data with factors when fitted through fit_xy() (#365).

  • proportional_hazards() with the "survival" engine now predicts when a single strata() term contains more than one variable, e.g. strata(s1, s2) (#404).

  • survival_prob_*() and hazard_*() now validate their inputs and return more informative error messages when given an unusable object, new_data, or eval_time (#271).

  • survival_reg() with the "survival" engine and strata() terms now returns NA survival and hazard predictions for new data rows with a missing value in a stratification variable, instead of silently using another stratum's scale (#383).

  • Engine-specific tuning parameters are now registered by censored instead of parsnip (#417).