Skip to contents

CRR Workflow

Source: vignettes/crr.Rmd
crr.Rmd

Read this article if you are fitting conditional risk ratios with hte_task() and fit_crr() and need to choose among the supported CRR learner families.

library(hte3)
library(sl3)

data <- hte3_example_data(n = 150, seed = 3)

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

crr_model <- fit_crr(
  task,
  method = "ipw",
  base_learner = Lrnr_mean$new(),
  cross_validate = FALSE
)

head(predict(crr_model, data))
#> [1] 0.1677935 0.1677935 0.1677935 0.1677935 0.1677935 0.1677935
summary(crr_model)
#> <summary.hte3_model>
#>   Target: CRR
#>   Engine: sl3
#>   Method: ipw
#>   Cross-validated: no
#>   Rows: 150
#>   Modifiers: W1, W2, W3
#>   Treatment variable: A

Choose the Wrapper Method

fit_crr() supports method = c("ep", "ipw", "t").

  • method = "ep" is the default flexible CRR path.
  • method = "ipw" is the weighting-based benchmark and places more emphasis on the propensity model.
  • method = "t" is the baseline built from arm-specific outcome models on the risk-ratio scale.

The wrapper pattern is the same as for CATE: build an hte3_Task, choose the CRR learner family with method, and optionally compare candidates with cross_validate = TRUE.

Compare CRR Learner Families 

cv_model <- fit_crr(
  task,
  method = c("ipw", "t", "ep"),
  base_learner = Lrnr_mean$new(),
  cross_validate = TRUE,
  cv_control = list(V = 2)
)

cv_model
#> <hte3_model>
#>   Target: CRR
#>   Engine: sl3
#>   Method: portfolio(ipw, t, ep)
#>   Cross-validated: yes
#>   Selected candidate: ep[basis=24, interaction=3]
#>   Selected method: ep
#>   Selected EP basis size: 24
#>   EP basis grid: 3, 6, 12, 18, 24
#>   Modifiers: W1, W2, W3
head(predict(cv_model, data))
#> [1] 0.151008 0.151008 0.151008 0.151008 0.151008 0.151008

If method = "ep" and cross_validate = TRUE, the wrapper also compares EP fits across basis sizes. Use sieve_basis_grid when you want to control that grid explicitly, and use sieve_num_basis for one fixed EP fit.

Important Constraints

  • CRR workflows require a non-negative outcome.
  • cross_fit in hte_task() controls nuisance estimation and is separate from wrapper-level learner selection.
  • For low-level CRR learner portfolios, see Advanced sl3 Integration.