Skip to contents

Advanced sl3 Integration

Source: vignettes/advanced-sl3.Rmd
advanced-sl3.Rmd

Read this article if you want the low-level API: explicit task construction, manual learner choice, or direct calls to cross_validate_cate() and cross_validate_crr().

Build the Task Explicitly

Use make_hte3_Task_tx() when you want to supply the nuisance learners yourself or work directly with the lower-level learner objects.

library(hte3)
library(sl3)

data <- hte3_example_data(n = 100, seed = 4)

task <- make_hte3_Task_tx(
  data = data,
  modifiers = c("W1", "W2"),
  confounders = c("W1", "W2", "W3"),
  treatment = "A",
  outcome = "Y",
  learner_pi = Lrnr_mean$new(),
  learner_mu = Lrnr_mean$new(),
  learner_m = Lrnr_mean$new(),
  cross_fit_and_cv = FALSE
)

Train a Learner Directly 

learner <- Lrnr_cate_DR$new(base_learner = Lrnr_mean$new())
trained <- learner$train(task)

head(predict_hte3(trained, data))
#> [1] 1.553732 1.553732 1.553732 1.553732 1.553732 1.553732

This is the right level when you want the exact learner object rather than the wrapper returned by fit_cate() or fit_crr().

Low-Level Cross-Validation

You can cross-validate low-level learner portfolios directly.

cate_cv <- cross_validate_cate(
  list(
    Lrnr_cate_DR$new(base_learner = Lrnr_mean$new()),
    Lrnr_cate_R$new(base_learner = Lrnr_mean$new()),
    Lrnr_cate_EP$new(base_learner = Lrnr_mean$new(), sieve_num_basis = 4)
  ),
  task,
  cv_control = list(V = 2)
)

cate_cv$coefficients
#>                                   dr                                    r 
#>                                    0                                    0 
#> ep[style=dr, basis=4, interaction=3] 
#>                                    1
head(cate_cv$lrnr_sl$predict(task))
#> [1] 1.547718 1.547718 1.547718 1.547718 1.547718 1.547718

The analogous CRR workflow is:

crr_task <- make_hte3_Task_tx(
  data = data,
  modifiers = c("W1", "W2", "W3"),
  confounders = c("W1", "W2", "W3"),
  treatment = "A",
  outcome = "Y_binary",
  learner_pi = Lrnr_mean$new(),
  learner_mu = Lrnr_mean$new(),
  learner_m = Lrnr_mean$new(),
  cross_fit_and_cv = FALSE
)

crr_cv <- suppressWarnings(
  cross_validate_crr(
    list(
      Lrnr_crr_IPW$new(base_learner = Lrnr_mean$new()),
      Lrnr_crr_T$new(base_learner = Lrnr_mean$new()),
      Lrnr_crr_EP$new(base_learner = Lrnr_mean$new(), sieve_num_basis = 4)
    ),
    crr_task,
    cv_control = list(V = 2)
  )
)

crr_cv$coefficients
#>                        ipw                          t 
#>                          0                          1 
#> ep[basis=4, interaction=3] 
#>                          0
head(crr_cv$lrnr_sl$predict(crr_task))
#> [1] 0.1634243 0.1634243 0.1634243 0.1634243 0.1634243 0.1634243

Important Constraints

  • cross_fit_and_cv in make_hte3_Task_tx() controls nuisance estimation at the task stage.
  • If modifiers = V and confounders = W with V a strict subset of W, the reduced-modifier target is E[Y(1) - Y(0) | V] = E[tau(W) | V].
  • In that reduced-modifier setting, the current low-level R-learner targets an overlap-weighted projection rather than the unweighted V-conditional CATE.
  • For continuous-treatment CATE, the low-level Lrnr_cate_R path is the one currently implemented.

For EP-specific sieve construction and tuning, see EP-Learner Sieve Tuning. If you are building a new learner, see Extending hte3 with Custom Learners.