Skip to contents

Causal forest with an isotonic monotonicity constraint on the CATE

Source: R/causal_forest_honest.R
morie_monotone_causal_forest.Rd

When theory says the effect can only move one way in a covariate, the unconstrained forest's wiggles are noise. The fitted CATE is projected onto the monotone functions of that covariate by pool-adjacent- violators, the exact L2 isotonic projection. Mirrors morie.fn.htgcrf.

Usage

morie_monotone_causal_forest(
  y,
  d,
  x,
  monotone_feature = NULL,
  direction = 1L,
  n_trees = 200L,
  min_leaf = 10L,
  seed = 0L
)

Arguments

y, d, x

As in morie_causal_forest().

monotone_feature

Column index the effect must be monotone in; NULL returns the unconstrained forest.

direction

1 for nondecreasing, -1 for nonincreasing.

n_trees, min_leaf, seed

Forest hyperparameters.

Value

List with cate, cate_raw, monotone_feature, direction, violations_before, violations_after, n.

References

Robertson T, Wright FT, Dykstra RL (1988). Order Restricted Statistical Inference. Wiley.

Examples

set.seed(4)
n <- 120
X <- matrix(rnorm(n * 2), n, 2)
d <- rbinom(n, 1, 0.5)
y <- 1 + X[, 1] + d * (0.5 + 0.5 * X[, 1]) + rnorm(n, 0, 0.3)
r <- morie_monotone_causal_forest(y, d, X, monotone_feature = 1L,
                                  n_trees = 20L)
str(r, max.level = 1)
#> List of 7
#>  $ cate             : num [1:120] 0.504 0.504 0.504 0.504 0.504 ...
#>  $ cate_raw         : num [1:120] 0.504 0.504 0.504 0.504 0.504 ...
#>  $ monotone_feature : int 1
#>  $ direction        : int 1
#>  $ violations_before: int 0
#>  $ violations_after : int 0
#>  $ n                : int 120