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.
# Install from GitHub
devtools::install_github("jaredhuling/forestBalance")# 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$Y)
fit
#> Forest Kernel Energy Balancing (cross-fitted)
#> --------------------------------------------------
#> n = 500 (n_treated = 156, n_control = 344)
#> Trees: 1000
#> Estimand: ATE
#> Cross-fitting: 2 folds
#> Solver: cg
#> ATE estimate: -0.0041
#> Fold ATEs: -0.2443, 0.2361
#> ESS: treated = 118/156 (76%) control = 283/344 (82%)
#> --------------------------------------------------
#> Use summary() for covariate balance details.The method proceeds in three steps:
-
Joint forest model: A
grf::multi_regression_forestis fit on covariates$X$ with the bivariate response$(A, Y)$ . Because the forest splits on both treatment and outcome, the resulting tree structure captures confounding relationships. -
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. -
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.
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 0.0000fit <- forest_balance(dat$X, dat$A, dat$Y, num.trees = 1000)print() gives a concise overview:
fit
#> Forest Kernel Energy Balancing (cross-fitted)
#> --------------------------------------------------
#> n = 800 (n_treated = 284, n_control = 516)
#> Trees: 1000
#> Estimand: ATE
#> Cross-fitting: 2 folds
#> Solver: direct
#> ATE estimate: 0.0986
#> Fold ATEs: 0.1494, 0.0477
#> ESS: treated = 209/284 (73%) control = 422/516 (82%)
#> --------------------------------------------------
#> Use summary() for covariate balance details.summary() provides a full covariate balance comparison (unweighted vs
weighted) with flagged imbalances:
summary(fit)
#> Forest Kernel Energy Balancing
#> ============================================================
#> n = 800 (n_treated = 284, n_control = 516)
#> Trees: 1000
#> Solver: CG (kernel not materialized)
#>
#> ATE estimate: 0.0986
#> ============================================================
#>
#> Covariate Balance (|SMD|)
#> ------------------------------------------------------------
#> Covariate Unweighted Weighted
#> ---------- ------------ ------------
#> X1 0.2467 * 0.0081
#> X2 0.0052 0.0235
#> X3 0.0690 0.0188
#> X4 0.0096 0.0023
#> X5 0.0382 0.0250
#> X6 0.0633 0.0200
#> X7 0.0354 0.0408
#> X8 0.0040 0.0211
#> X9 0.1140 * 0.0162
#> X10 0.1740 * 0.1261 *
#> ---------- ------------ ------------
#> Max |SMD| 0.2467 0.1261
#> (* indicates |SMD| > 0.10)
#>
#> Effective Sample Size
#> ------------------------------------------------------------
#> Treated: 209 / 284 (73%)
#> Control: 422 / 516 (82%)
#>
#> Energy Distance
#> ------------------------------------------------------------
#> Unweighted: 0.0397
#> Weighted: 0.0173
#> ============================================================Since confounding operates through nonlinear functions of
X <- dat$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
#> ============================================================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 0.2063For 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$A == 0])
c("ATE estimate" = round(ate, 4))
#> ATE estimate
#> 0.3684A small simulation comparing the forest balance estimator against the naive (unadjusted) difference in means:
set.seed(1)
nreps <- 100
results <- matrix(NA, nreps, 2, dimnames = list(NULL, c("Naive", "Forest")))
for (r in seq_len(nreps)) {
dat <- simulate_data(n = 500, p = 10, ate = 0)
fit <- forest_balance(dat$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 |
| 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 |
