Skip to content

About

Balancing Confounder Distributions with Forest Energy Balancing

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

forestBalance

Forest Kernel Energy Balancing for Causal Inference

forestBalance estimates average treatment effects (ATE) by combining multivariate random forests with kernel energy balancing. A joint forest model of covariates, treatment, and outcome defines a proximity kernel that characterizes the confounding structure and emphasizes similarity of observations in terms of confounding. The kernel is then used in kernel energy balancing, which minimizes the distributional distance between treatment and control groups. By using the joint forest kernel, this balancing approach emphasizes balance of variables that impact both treatment and outcome. By construction, these balancing weights aim to balance the joint distribution of confounders specifically.

The method is described in:

De, S. and Huling, J.D. (2025). Data adaptive covariate balancing for causal effect estimation for high dimensional data. arXiv:2512.18069.

Installation

# Install from GitHub
devtools::install_github("jaredhuling/forestBalance")

Quick start

# library(forestBalance)

# Simulate observational data with nonlinear confounding (true ATE = 0)
set.seed(123)
dat <- simulate_data(n = 500, p = 10, ate = 0)

# Estimate ATE with forest kernel energy balancing
fit <- forest_balance(dat$X, dat$A, dat$Y)
fit
#> Forest Kernel Energy Balancing (cross-fitted)
#> -------------------------------------------------- 
#>   n = 500  (n_treated = 156, n_control = 344)
#>   Trees: 1000
#>   Estimand: ATE
#>   Cross-fitting: 2 folds
#>   Solver: cg
#>   ATE estimate: -0.0041
#>   Fold ATEs: -0.2443, 0.2361
#>   ESS: treated = 118/156 (76%)   control = 283/344 (82%)
#> -------------------------------------------------- 
#> Use summary() for covariate balance details.

How it works

The method proceeds in three steps:

  1. Joint forest model: A grf::multi_regression_forest is fit on covariates $X$ with the bivariate response $(A, Y)$. Because the forest splits on both treatment and outcome, the resulting tree structure captures confounding relationships.

  2. Proximity kernel: The $n \times n$ kernel matrix $K(i,j)$ is defined as the proportion of trees where observations $i$ and $j$ share a leaf node. This is computed efficiently via a single sparse matrix cross-product.

  3. Kernel energy balancing: Balancing weights are obtained in closed form by solving a linear system derived from the kernel energy distance objective. The weights make the treated and control distributions similar with respect to the forest-defined similarity measure.

Detailed example

Simulating data

simulate_data() generates observational data with nonlinear confounding through a Beta density link:

set.seed(123)
dat <- simulate_data(n = 800, p = 10, ate = 0)

# Naive (unweighted) estimate is biased
naive_ate <- mean(dat$Y[dat$A == 1]) - mean(dat$Y[dat$A == 0])
c("Naive ATE" = round(naive_ate, 4), "True ATE" = 0)
#> Naive ATE  True ATE 
#>    1.7054    0.0000

Fitting the model

fit <- forest_balance(dat$X, dat$A, dat$Y, num.trees = 1000)

Print and summary

print() gives a concise overview:

fit
#> Forest Kernel Energy Balancing (cross-fitted)
#> -------------------------------------------------- 
#>   n = 800  (n_treated = 284, n_control = 516)
#>   Trees: 1000
#>   Estimand: ATE
#>   Cross-fitting: 2 folds
#>   Solver: direct
#>   ATE estimate: 0.0986
#>   Fold ATEs: 0.1494, 0.0477
#>   ESS: treated = 209/284 (73%)   control = 422/516 (82%)
#> -------------------------------------------------- 
#> Use summary() for covariate balance details.

summary() provides a full covariate balance comparison (unweighted vs weighted) with flagged imbalances:

summary(fit)
#> Forest Kernel Energy Balancing
#> ============================================================ 
#>   n = 800  (n_treated = 284, n_control = 516)
#>   Trees: 1000
#>   Solver: CG (kernel not materialized)
#> 
#>   ATE estimate: 0.0986
#> ============================================================ 
#> 
#> Covariate Balance (|SMD|)
#> ------------------------------------------------------------ 
#>   Covariate     Unweighted      Weighted
#>   ----------  ------------  ------------
#>   X1              0.2467 *        0.0081
#>   X2                0.0052        0.0235
#>   X3                0.0690        0.0188
#>   X4                0.0096        0.0023
#>   X5                0.0382        0.0250
#>   X6                0.0633        0.0200
#>   X7                0.0354        0.0408
#>   X8                0.0040        0.0211
#>   X9              0.1140 *        0.0162
#>   X10             0.1740 *      0.1261 *
#>   ----------  ------------  ------------
#>   Max |SMD|         0.2467        0.1261
#>   (* indicates |SMD| > 0.10)
#> 
#> Effective Sample Size
#> ------------------------------------------------------------ 
#>   Treated: 209 / 284  (73%)
#>   Control: 422 / 516  (82%)
#> 
#> Energy Distance
#> ------------------------------------------------------------ 
#>   Unweighted: 0.0397
#>   Weighted:   0.0173
#> ============================================================

Balance on nonlinear transformations

Since confounding operates through nonlinear functions of $X_1$ and $X_5$, we can check balance on transformations of the covariates:

X <- dat$X
X.nl <- cbind(
  X[,1]^2, X[,2]^2, X[,5]^2,
  X[,1] * X[,2], X[,1] * X[,5],
  dbeta(X[,1], 2, 4), dbeta(X[,5], 2, 4)
)
colnames(X.nl) <- c("X1^2", "X2^2", "X5^2", "X1*X2", "X1*X5",
                     "Beta(X1)", "Beta(X5)")

summary(fit, X.trans = X.nl)
#> Forest Kernel Energy Balancing
#> ============================================================ 
#>   n = 800  (n_treated = 284, n_control = 516)
#>   Trees: 1000
#>   Solver: CG (kernel not materialized)
#> 
#>   ATE estimate: 0.0986
#> ============================================================ 
#> 
#> Covariate Balance (|SMD|)
#> ------------------------------------------------------------ 
#>   Covariate     Unweighted      Weighted
#>   ----------  ------------  ------------
#>   X1              0.2467 *        0.0081
#>   X2                0.0052        0.0235
#>   X3                0.0690        0.0188
#>   X4                0.0096        0.0023
#>   X5                0.0382        0.0250
#>   X6                0.0633        0.0200
#>   X7                0.0354        0.0408
#>   X8                0.0040        0.0211
#>   X9              0.1140 *        0.0162
#>   X10             0.1740 *      0.1261 *
#>   ----------  ------------  ------------
#>   Max |SMD|         0.2467        0.1261
#>   (* indicates |SMD| > 0.10)
#> 
#> Transformed Covariate Balance (|SMD|)
#> ------------------------------------------------------------ 
#>   Transform     Unweighted      Weighted
#>   ----------  ------------  ------------
#>   X1^2            0.1769 *        0.0394
#>   X2^2              0.0125        0.0079
#>   X5^2              0.0453        0.0474
#>   X1*X2             0.0397        0.0176
#>   X1*X5             0.0407        0.0025
#>   Beta(X1)        0.7025 *        0.0353
#>   Beta(X5)        0.1256 *        0.0165
#>   ----------  ------------  ------------
#>   Max |SMD|         0.7025        0.0474
#> 
#> Effective Sample Size
#> ------------------------------------------------------------ 
#>   Treated: 209 / 284  (73%)
#>   Control: 422 / 516  (82%)
#> 
#> Energy Distance
#> ------------------------------------------------------------ 
#>   Unweighted: 0.0397
#>   Weighted:   0.0173
#> ============================================================

Standalone balance diagnostics

compute_balance() can be used independently with any set of weights:

# Inverse propensity weights (using true propensity scores)
ipw <- ifelse(dat$A == 1, 1 / dat$propensity, 1 / (1 - dat$propensity))

bal_forest <- compute_balance(dat$X, dat$A, fit$weights)
bal_ipw    <- compute_balance(dat$X, dat$A, ipw)

c("Forest balance" = round(bal_forest$max_smd, 4),
  "IPW"            = round(bal_ipw$max_smd, 4))
#> Forest balance            IPW 
#>         0.1261         0.2063

Lower-level interface

For more control, the pipeline can be run step by step:

library(grf)

# 1. Fit the joint forest
forest <- multi_regression_forest(dat$X, scale(cbind(dat$A, dat$Y)),
                                  num.trees = 500)

# 2. Extract leaf node matrix and build kernel
leaf_mat <- get_leaf_node_matrix(forest, dat$X)
K <- leaf_node_kernel(leaf_mat)

c("observations" = nrow(leaf_mat), "trees" = ncol(leaf_mat))
#> observations        trees 
#>          800          500
c("kernel % nonzero" = round(100 * length(K@x) / prod(dim(K)), 1))
#> kernel % nonzero 
#>             24.3

# 3. Compute balancing weights
bal <- kernel_balance(dat$A, K)

ate <- weighted.mean(dat$Y[dat$A == 1], bal$weights[dat$A == 1]) -
       weighted.mean(dat$Y[dat$A == 0], bal$weights[dat$A == 0])
c("ATE estimate" = round(ate, 4))
#> ATE estimate 
#>       0.3684

Simulation study

A small simulation comparing the forest balance estimator against the naive (unadjusted) difference in means:

set.seed(1)
nreps <- 100
results <- matrix(NA, nreps, 2, dimnames = list(NULL, c("Naive", "Forest")))

for (r in seq_len(nreps)) {
  dat <- simulate_data(n = 500, p = 10, ate = 0)
  fit <- forest_balance(dat$X, dat$A, dat$Y, num.trees = 500)
  results[r, "Naive"]  <- mean(dat$Y[dat$A == 1]) - mean(dat$Y[dat$A == 0])
  results[r, "Forest"] <- fit$ate
}
Method Bias SD RMSE
Naive 1.2944 0.2984 1.3284
Forest 0.1086 0.1613 0.1945

Key functions

Function Description
forest_balance() High-level: fit forest, build kernel, compute weights, return ATE
simulate_data() Simulate observational data with nonlinear confounding
compute_balance() Covariate balance diagnostics (SMD, ESS, energy distance)
get_leaf_node_matrix() Fast vectorized leaf node extraction from grf forests
leaf_node_kernel() Sparse proximity kernel from leaf node matrix
forest_kernel() Convenience: forest object to kernel in one call
kernel_balance() Closed-form kernel energy balancing weights

About

Balancing Confounder Distributions with Forest Energy Balancing

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages