Skip to contents

Quickstart

Source: vignettes/quickstart.Rmd
quickstart.Rmd

Read this article if you want one minimal end-to-end example with the high-level wrappers before choosing a more specific workflow.

library(hte3)
library(sl3)

data <- hte3_example_data(n = 120, seed = 1)

task <- hte_task(
  data = data,
  modifiers = c("W1", "W2", "W3"),
  confounders = c("W1", "W2", "W3"),
  treatment = "A",
  outcome = "Y",
  propensity_learner = Lrnr_mean$new(),
  outcome_learner = Lrnr_mean$new(),
  mean_learner = Lrnr_mean$new(),
  cross_fit = FALSE
)

model <- fit_cate(
  task,
  method = "dr",
  base_learner = Lrnr_mean$new(),
  cross_validate = FALSE
)

head(predict(model, data))
#> [1] 1.027376 1.027376 1.027376 1.027376 1.027376 1.027376
summary(model)
#> <summary.hte3_model>
#>   Target: CATE
#>   Engine: sl3
#>   Method: dr
#>   Cross-validated: no
#>   Rows: 120
#>   Modifiers: W1, W2, W3
#>   Treatment variable: A

Cross-Fit vs Cross-Validate

Two arguments control different stages of the workflow:

  • cross_fit in hte_task() controls nuisance estimation.
  • cross_validate in fit_cate() or fit_crr() controls selection across candidate HTE learners.

These are separate decisions. A common analysis cross-fits nuisances and only turns on wrapper-level cross-validation when comparing learner families.