propensity_learner
Estimates treatment assignment. This matters especially for DR, R, EP, and IPW style workflows.
Learner Slots
Where do learners plug into hte3?
outcome_learner
Estimates the outcome regression. This is often one of the main quality drivers for CATE and CRR performance.
mean_learner
Provides the marginal mean nuisance when the workflow requires it. In many analyses, the package default is sufficient.
base_learner
Used inside the chosen meta-learner to learn the final heterogeneity surface after the pseudo-outcome step.
Default Stack
What does get_autoML() use?
Stack$new(
Lrnr_glmnet$new(),
Lrnr_gam$new(),
# Added when optional runtime packages are installed:
Lrnr_earth$new(degree = 2),
Lrnr_ranger$new(max.depth = 10),
Lrnr_xgboost_early_stopping$new(min_child_weight = 15, max_depth = 2, eta = 0.2, subsample = 0.8, colsample_bytree = 0.8),
Lrnr_xgboost_early_stopping$new(min_child_weight = 15, max_depth = 3, eta = 0.15, subsample = 0.9),
Lrnr_xgboost_early_stopping$new(min_child_weight = 15, max_depth = 4, eta = 0.15, subsample = 0.9),
Lrnr_xgboost_early_stopping$new(min_child_weight = 15, max_depth = 5, eta = 0.15, subsample = 0.9),
Lrnr_xgboost_early_stopping$new(min_child_weight = 15, max_depth = 4, eta = 0.08, subsample = 0.8, colsample_bytree = 0.8),
)Why this is useful
It mixes linear, additive, spline, tree, and boosting-style learners, which gives you a reasonable first-pass library without designing one from scratch.
When to override it
Override the default when you need shorter iteration cycles, stronger interpretability, or a task-specific learner library.
One caveat
Some sl3 learners depend on supporting packages being available in your R environment,
so actual availability can vary by setup.
Stacking
How Stack and Lrnr_sl work in practice
In sl3, a Stack is just a library of candidate learners.
A Super Learner is what you get when you wrap that library in
Lrnr_sl with a metalearner that combines or selects among the candidates.
Build a learner library
Use Stack$new(...) to enumerate the models you want to compare.
This does not fit an ensemble by itself. It just defines the candidate set.
library(sl3)
learner_library <- Stack$new(
Lrnr_glm_fast$new(),
Lrnr_ranger$new(),
Lrnr_xgboost$new()
)Wrap it in a Super Learner
Use Lrnr_sl$new(...) when you want cross-validated selection or
weighted combination across that library.
sl_fit <- Lrnr_sl$new(
learners = learner_library,
metalearner = Lrnr_nnls$new(),
cv_control = list(V = 5)
)$train(task)What Stack means in hte3
You can pass a stacked library anywhere hte3 expects a learner, such as
propensity_learner, outcome_learner, or base_learner.
What Lrnr_sl adds
Lrnr_sl adds outer learner selection or weighted combination on top of a
library, using a metalearner and fold structure defined through cv_control.
Common hte3 default
get_autoML() returns a Stack built from its available safe
default learners, which gives you a ready-made candidate library without manually
listing learners.
Practical rule:
use
Stack when you want to define the candidates, and use
Lrnr_sl when you want sl3 to cross-validate and combine them.
Cross-Validation
There are three different CV layers to keep straight
Users often say “cross-validation” when they mean different parts of the pipeline.
In hte3, it helps to separate nuisance cross-fitting, sl3
learner-library CV, and outer HTE learner selection.
1. Nuisance cross-fitting
Controlled by cross_fit = TRUE in hte_task() or
cross_fit_and_cv = TRUE in the low-level task builder.
This is about estimating nuisance functions more robustly, not selecting among HTE methods.
task <- hte_task(
data = df,
modifiers = mods,
confounders = confs,
treatment = "A",
outcome = "Y",
cross_fit = TRUE
)2. sl3 library CV
Controlled inside sl3 with Lrnr_sl and
cv_control = list(V = ...). This is where a Super Learner compares
members of a Stack.
sl_fit <- Lrnr_sl$new(
learners = learner_library,
metalearner = Lrnr_nnls$new(),
cv_control = list(V = 5)
)$train(task)3. Outer HTE learner CV
Controlled by cross_validate = TRUE in fit_cate() or
fit_crr(), or explicitly with cross_validate_cate() and
cross_validate_crr(). This is where you compare DR, R, T, EP, or CRR families.
fit <- fit_cate(
task,
method = c("dr", "r", "ep"),
base_learner = learner_library,
cross_validate = TRUE,
cv_control = list(V = 5)
)
Recommended mental model:
cross_fit is for nuisance estimation, Lrnr_sl is for
sl3 Super Learner selection inside a learner library, and
cross_validate = TRUE in the wrappers is for selection among HTE learner families.
Starter Recipes
Common learner strategies for practitioners
Simple generalized linear learners
Use learners like Lrnr_glm_fast or Lrnr_mean for rapid iteration,
debugging, or an interpretable baseline.
Keep get_autoML()
Use this when a broad learner library is needed without constructing one manually.
Tree and boosting heavy
Favor learners such as Lrnr_ranger and Lrnr_xgboost when interactions and nonlinearities are likely to matter.
Hand-built Stack or Lrnr_sl
Use this path when a preferred learner library already exists or when comparing a curated set of candidate learners.
Code Patterns
Small patterns you can copy into real analyses
Fast and explicit
hte_task(
data = df,
modifiers = mods,
confounders = confs,
treatment = "A",
outcome = "Y",
propensity_learner = Lrnr_glm_fast$new(),
outcome_learner = Lrnr_glm_fast$new(),
mean_learner = Lrnr_mean$new()
)Custom base learner for CATE
fit_cate(
task,
method = "dr",
base_learner = Lrnr_ranger$new(),
cross_validate = FALSE
)
Continuous-treatment note:
in the current package, the continuous-treatment CATE path is the
R-learner. That implementation uses the partially linear effect-model
view with an
A * tau(X) term rather than a fully general
treatment-response surface.
Reduced-modifier-set note:
if the target modifiers are
V and the nuisance adjustment
set is W with V a strict subset of W,
then the target of interest is E[Y(1)-Y(0) | V] = E[tau(W) | V].
In the supported binary/categorical-treatment setting, DR- and EP-learners
target that surface. The current R-learner instead targets the
overlap-weighted projection
f_R(V) = E[Var(A|W) tau(W) | V] / E[Var(A|W) | V],
which becomes
E[e(W)(1-e(W)) tau(W) | V] / E[e(W)(1-e(W)) | V]
for binary treatment.
Official sl3 Resources
References for further detail
These are the upstream references most likely to help an hte3 user make better modeling decisions.
sl3 package site
The main package homepage with reference docs and articles.
Intro to sl3
A practical article for understanding tasks, learners, training, and prediction.
Pipeline reference
Useful for chained learners, screening steps, or more customized workflows.
Super Learner reference
Use this reference when building or interpreting Lrnr_sl libraries explicitly.
tlverse handbook chapter
A longer-form chapter that is often easier to learn from than jumping straight into reference pages.