What does min.node.size actually do?

[tcovert]

Should I expect that if I set min.node.size = 10, every leaf of every tree in a causal_forest should have at least 10 sample's? Or am misunderstanding what min.node.size is supposed to mean? I have found that setting positive values for min.node.size does not result in leaves with at least that many samples.

Here is an MWE inspired by the examples in the documentation:

library(grf)
library(purrr)

leafsamples <- function(t) {
  leaves <-
    t %>%
    pluck("nodes") %>%
    keep("is_leaf") %>%
    map(pluck("samples"))
  return(leaves)
}

n = 2000; p = 10
X = matrix(rnorm(n*p), n, p)
W = rbinom(n, 1, 0.5)
Y = pmax(X[,1], 0) * W + X[,2] + pmin(X[,3], 0) + rnorm(n)
tau.forest = causal_forest(X, Y, W, min.node.size = 4, seed = 1)

leaves <- leafsamples(get_tree(tau.forest, 1))

On my machine, only 53 of the 159 leaves in the first tree of this forest have 4 or more samples.

[aliciaMAR]

My understanding was that min.node.size should constraint the minimum nb of units from treatment (resp. control) in terminal nodes, for a given tree. But indeed, as tcovert, causal_forest with min.node.size = 10L returns trees with terminal nodes containing sometimes only 1 sample !

[jtibshirani]

Currently, we don't prevent splits from occurring that could result in nodes with size less than min.node.size. Instead, the algorithm simply stops splitting if a node's size is less than or equal to min.node.size. Our core splitting implementation is based on the ranger package (which is in turn based on Breiman + Cutler's randomForest package). Both packages make this approximation around min.node.size.

I agree that this behavior is quite misleading, and I've filed #143 to track the issue. For now, I'll add documentation to explain why there is a discrepancy in node sizes.