Native stacking: fits a small library of propensity learners
(logistic, logistic with squared terms, and rmorie's native random
forest – a probability forest of 200 CART trees, mtry =
\(\lfloor\sqrt p\rfloor\), minimum node size 10, the ranger
probability-forest defaults – so the library is identical on every
install), combines their cross-validated predictions by
non-negative least squares on the CV log-loss surface (native NNLS
via Lawson-Hanson active set), and converts the ensemble scores to
estimand weights.
Usage
morie_weight_super(data, treatment, covariates, estimand = "ATE", n_folds = 5L)
Arguments
- data
Data frame.
- treatment
Binary treatment column (0/1).
- covariates
Covariate column names.
- estimand
"ATE" (default), "ATT", or "ATC".
- n_folds
Cross-validation folds for the stacking weights.
Value
A morie_weight object; $learner_weights gives
the ensemble coefficients.
References
van der Laan, Polley & Hubbard (2007).
Examples
set.seed(1)
d <- data.frame(t = rbinom(150, 1, 0.5), x1 = rnorm(150), x2 = rnorm(150))
morie_weight_super(d, "t", c("x1", "x2"))
#> morie_weight: SuperLearner (native NNLS stack) (estimand ATE)
#> n = 150 ESS = 143.2 weight range [1.129, 4.048]
