Finding Subgroups in Treatment Effects with Conformal Trees

Usage

r2p_hte(
  data,
  target,
  treatment,
  learner,
  cv_folds = 1,
  alpha = 0.1,
  gamma = 0.01,
  lambda = 0.5,
  max_groups = 5
)

Arguments

data: (data.frame)
data set for model training and uncertainty estimation.
target: (string)
name of the target variable. The target must be a numeric variable.
treatment: (string)
name of the treatment variable. If treatment
is a factor, then the first level is treated as control and the second level as treatment indicator. If treatment
is a numeric, then zero-one encoding is assumed and "1"
treated as treatment indicator.
learner: (model_spec)
the learner for training the prediction model. See parsnip::model_spec() for details.
cv_folds: (count)
number of CV+ folds.
alpha: (proportion)
miscoverage rate.
gamma: (proportion)
regularization parameter ensuring that reduction in the impurity of the confident homogeneity is sufficiently large.
lambda: (proportion)
balance parameter, quantifying the impact of the average interval length relative to the average absolute deviation (i.e. interval width vs. average absolute deviation)
max_groups: (count)
maximum number of subgroups.

Value

The tree.

Examples

library(tidymodels)
# Synthetic example data:
library(htesim)
set.seed(12)
dgp <- dgp(p = pF_exp_x1_x2,
          m = mF_x1,
          t = tF_div_x1_x2,
          model = "normal",
          xmodel = "unif",
          sd = 1)
sim <- simulate(object = dgp,
               nsim = 500L,
               d = 4L)
# Initialize learner:
linear <- linear_reg() %>%
 set_mode("regression") %>%
 set_engine("lm")
# Detect subgroups:
groups <- r2p_hte(
  data = sim,
  target = "y",
  treatment = "trt",
  learner = linear,
  cv_folds = 500,
  alpha = 0.1,
  gamma = 0.01,
  lambda = 0.5,
  max_groups = 8
)
summary(groups)
#> Conformal tree with 3 subgroups:
#>     n mean width deviation
#> 1 258 0.25  8.47         0
#> 2  79 0.60  6.80         0
#> 3 163 0.66  9.29         0
#> ---
#> Alpha:  0.1 Lambda:  0.5 Gamma:  0.01