ensembleHTE

Overview

ensembleHTE detects and characterizes heterogeneous treatment effects (HTE) in randomized experiments, and supports standard prediction tasks with valid downstream inference. It uses repeated cross-fitting—where every observation is used for both training and evaluation—and combines predictions from multiple ML algorithms to produce stable estimates and avoid the multiple-testing issues that arise from having to pick a single best algorithm.

Two main entry points:

• ensemble_hte() — estimate heterogeneous treatment effects.
• ensemble_pred() — predict an outcome (no treatment structure needed).

A suite of analysis functions—blp(), gates(), gavs(), clan(), and their restricted counterparts—lets you test for heterogeneity, characterize who is in each group, and compare targeting strategies.

For the statistical foundations, see Fava (2025).

Installation

You can install the development version of ensembleHTE from GitHub:

# install.packages("devtools")
devtools::install_github("bfava/ensembleHTE")

Quick Start

library(ensembleHTE)
data(microcredit)

# 1. Fit ensemble HTE model
covariates <- c("css_creditscorefinal", "lower_window",
                "own_anybus", "max_yearsinbusiness", "css_assetvalue")

fit <- ensemble_hte(
  Y = "exp_yrly_end",
  D = "treat",
  X = covariates,
  data       = microcredit,
  prop_score = "prop_score",
  algorithms = c("grf", "glmnet", "xgboost"),
  M = 20,
  K = 4
)

# 2. View results
print(fit)
summary(fit)

# 3. Analyze treatment effect heterogeneity
blp_results   <- blp(fit)                    # Best Linear Predictor
gates_results <- gates(fit, n_groups = 3)    # Group Average Treatment Effects
clan_results  <- clan(fit, n_groups = 3)     # Classification Analysis

# 4. Visualize
plot(gates_results)
plot(clan_results)

# --- Prediction task (no treatment structure) ---

# Predict bank profits (observed only for borrowers who took a loan)
has_loan <- microcredit$treat == 1 & microcredit$loan_size > 0

fit_pred <- ensemble_pred(
  Y    = "bank_profits_pp",
  X    = covariates,
  data = microcredit,
  train_idx  = has_loan,
  algorithms = c("grf", "glmnet", "xgboost"),
  M = 20,
  K = 4
)

summary(fit_pred)
gavs_results <- gavs(fit_pred, n_groups = 3)
plot(gavs_results)

Main Functions

Estimation

• ensemble_hte(): Fit ensemble heterogeneous treatment effect model
• ensemble_pred(): Fit ensemble prediction model (without treatment effects)

Analysis

• blp() / blp_pred(): Best Linear Predictor (test for heterogeneity / calibration)
• gates(): Group Average Treatment Effects
• gavs(): Group Averages for prediction tasks
• clan(): Classification Analysis — characterize high/low effect groups

Restricted comparisons

• gates_restricted(): Compare GATES under unrestricted vs. equity-constrained targeting
• gavs_restricted(): Compare GAVS under unrestricted vs. equity-constrained targeting

Getting Help

If you encounter a bug or have a feature request, please file an issue.

Citation

If you use this package in your research, please cite:

@article{fava2025training,
  title={Training and Testing with Multiple Splits: A Central Limit Theorem for Split-Sample Estimators},
  author={Fava, Bruno},
  journal={arXiv preprint arXiv:2511.04957},
  year={2025}
}

License

Apache License 2.0 — see LICENSE file for details