wpor: a package for weighted pseudo-outcome regression

This package implements the meta-learner methods discussed in Fisher (2023), including

Inverse Variance Weighted Doubly Robust Learning,
Standard Doubly Robust (DR) Learning,
Robinson (R) Learning,
U Learning,
Covariance-based Learning,
Inverse Variance Weighted, Covariance-based Learning, and
Two-Model (T) Learning.

Installation

devtools::install_github('aaronjfisher/wpor')

Specifying the nuisance models and the conditional effect model

As WPOR is a meta algorithm that is agnostic to the models used, we aimed to make wpor compatible with a wide variety of chocies. Rather than hard-coding a handful of options for nuisance models, wpor integrates with the collection of training algorithms in the tidymodels set of packages, namely parsnip and workflows. These packages provide a standardized syntax for many popular training algorithms. The simplest way to specify a training algorithm in wpor is to define a "workflow" object (see examples in the "Fitting WPOR Models" vignette).

For examples of how to extend wpor to use custom algorithms not available in parsnip, see the "Custom models" vignette.

Parameter tuning

In general when implementing cross-fitting with hyper-parameter tuning, there are two common approaches:

Strict cross-fitting with nested tuning: within each iteration of cross-fitting, run cross-validation to select tuning parameters.
Approximate cross-fitting with pre-tuned parameters: Before running cross-fitting, pre-tune the parameters with cross-validation. Next, use the same tuning parameters over all iterations of cross-fitting. This technically breaks the independencies that cross-fitting aims to ensure, as all data folds are used in selecting tuning parameters. At the time of this writing, this approximate approach is what is used in the rlearner package.

wpor includes convenience functions for either approach. For the approximate method, see ?tune_params. For the strict method, we suggest using the object class tunefit, which packages a workflow or training algorithm together with arguments for how that algorithm should be tuned (see ?as.tunefit). These instructions are then applied within in each iteration of cross-fitting.