NEWS.md
Add methods to configs which allow the addition of models and moderators after the initial instantiation. This is a first step towards the eventual recipe API.
If no moderators are specified in call to estimate_QoI, then all moderators listed in the MCATE definition are used.
First version of the recipe API is available for use!
Move joint effect model config to a single unified location in HTE_cfg.
Save the names of features used in SuperLearner ensembles so that when out-of-sample predictions are requested, columns of all zeros can be imputed for features that do not exist. This imputation will simply eliminate errors when there is a factor variable with a level observed in training but not in testing. This error would never appear in the case of a continuous covariate, because it will always exist across splits, so it only eliminates errors due to one-hot-encoding factor variables.
Add an initial draft of a vignette on methodological details undergirding the package.
Adds a linear-only version of variable importance that calculates the reduction in residual MSE from including a particular moderator.
Adds a Regression ROC Curve diagnostic for regression models.
Allows access to predictions from effect regression.
Fixes to VIMP. Arguments to cv_vim were previously constructed incorrectly.
Allows users to specify whether VIMP uses sample splitting (and is therefore has inference robust to zero variable importance) or not (lower variance of estimates). The default is to use sample splitting.
Export attach_config.
Clean up vignette a little.
Add support for different pseudo-outcomes, for example based on an IPW (only), or direct-estimation (only) in addition to the (default) doubly-robust method.
glmnet as a model in the VIMP regression by setting the (unused) censoring model to be SL.mean (a constant model, which is fast and easy to estimate).Increment to first minor version number. The package is generally feature-complete enough for general usage, although there are still lots of things to do before an initial major version release.
Public API no longer constantly passes in the HTE_cfg. Instead, a new public function is added, attach_config, which adds metadata to a given dataframe expressing how HTEs should be calculated.
Listwise deletion no longer fully drops rows from the dataset. It instead will drop rows at each step based on the columns that are necessary to be non-missing in that particular step.
soft_require now uses the underlying rlang::check_installed to also prompt the end-user to install necessary packages.
Adds support for population weights. This is a large change affecting a lot of functionality.
Discrete MCATEs now support population weights.
Continuous MCATEs will thrown an error indicating that nprobust does not support weights, currently. A subsequent release will provide other ways to smooth a final estimate with, e.g., binscatter methods.
calculate_ate now returns both a SATE and (when weights are given) a PATE.
Diagnostics now support weighting and will return the population version of themselves (e.g. AUC / MSE). This entailed switching AUC calculation from the pROC package to WeightedROC.
An additional suite of tests have been added to check that weighted analyses work correctly and have at least some modicum of correctness.
Add tests for check_identifier, splitting multiple times on a dataset, soft_require, output of SLPredictor.
Test for correctness of clustered standard error estimates (against estimatr::lm_robust).
Force progress bar to print even if the stream doesn’t seem to support it. This may cause some log-spew when used in batch mode, but will at least be informative about the place in the fitting process.
Add tests for VIMP, listwise_deletion
Add validity checking for the identifier column.
Fix R CMD check issue with VIMP not having SuperLearner attached.
Fix VIMP issue from using make_splits after a .split_id column already exists in the data.
Add sorting to API Reference web page
Add NEWS.md
A small selection of stylistic changes to remove lint messages.
Bandwidth selection in kernel smoothing for PCATEs now uses the appropriate evaluation points.
Adds a message to the user when a lack of quickblock installed causes reversion to un-stratified CV.