grf: generalized random forests

A pluggable package for forest-based statistical estimation and inference. GRF currently provides non-parametric methods for least-squares regression, quantile regression, and treatment effect estimation (optionally using instrumental variables).

In addition, GRF supports 'honest' estimation (where one subset of the data is used for choosing splits, and another for populating the leaves of the tree), and confidence intervals for least-squares regression and treatment effect estimation.

This package is currently in beta, and we expect to make continual improvements to its performance and usability. For a practical description of the GRF algorithm, including explanations of model parameters and troubleshooting suggestions, please see the GRF reference.

Authors

This package is written and maintained by Julie Tibshirani (jtibs@cs.stanford.edu), Susan Athey, and Stefan Wager.

The repository first started as a fork of the ranger repository -- we owe a great deal of thanks to the ranger authors for their useful and free package.

Installation

The latest release of the package can be installed through CRAN:

install.packages("grf")

Any published release can also be installed from source:

install.packages("https://raw.github.com/grf-labs/grf/master/releases/grf_0.10.2.tar.gz", repos = NULL, type = "source")

conda users can install from the conda-forge channel:

conda install -c conda-forge r-grf

Note that to install from source, a compiler that implements C++11 is required (clang 3.3 or higher, or g++ 4.8 or higher). If installing on Windows, the RTools toolchain is also required.

Usage Examples

The following script demonstrates how to use GRF for heterogeneous treatment effect estimation. For examples of how to use types of forest, as for quantile regression and causal effect estimation using instrumental variables, please consult the R documentation on the relevant forest methods (quantile_forest, instrumental_forest, etc.).

library(grf)

# Generate data.
n = 2000; p = 10
X = matrix(rnorm(n*p), n, p)
X.test = matrix(0, 101, p)
X.test[,1] = seq(-2, 2, length.out = 101)

# Train a causal forest.
W = rbinom(n, 1, 0.4 + 0.2 * (X[,1] > 0))
Y = pmax(X[,1], 0) * W + X[,2] + pmin(X[,3], 0) + rnorm(n)
tau.forest = causal_forest(X, Y, W)

# Estimate treatment effects for the training data using out-of-bag prediction.
tau.hat.oob = predict(tau.forest)
hist(tau.hat.oob$predictions)

# Estimate treatment effects for the test sample.
tau.hat = predict(tau.forest, X.test)
plot(X.test[,1], tau.hat$predictions, ylim = range(tau.hat$predictions, 0, 2), xlab = "x", ylab = "tau", type = "l")
lines(X.test[,1], pmax(0, X.test[,1]), col = 2, lty = 2)

# Estimate the conditional average treatment effect on the full sample (CATE).
average_treatment_effect(tau.forest, target.sample = "all")

# Estimate the conditional average treatment effect on the treated sample (CATT).
# Here, we don't expect much difference between the CATE and the CATT, since
# treatment assignment was randomized.
average_treatment_effect(tau.forest, target.sample = "treated")

# Add confidence intervals for heterogeneous treatment effects; growing more trees is now recommended.
tau.forest = causal_forest(X, Y, W, num.trees = 4000)
tau.hat = predict(tau.forest, X.test, estimate.variance = TRUE)
sigma.hat = sqrt(tau.hat$variance.estimates)
plot(X.test[,1], tau.hat$predictions, ylim = range(tau.hat$predictions + 1.96 * sigma.hat, tau.hat$predictions - 1.96 * sigma.hat, 0, 2), xlab = "x", ylab = "tau", type = "l")
lines(X.test[,1], tau.hat$predictions + 1.96 * sigma.hat, col = 1, lty = 2)
lines(X.test[,1], tau.hat$predictions - 1.96 * sigma.hat, col = 1, lty = 2)
lines(X.test[,1], pmax(0, X.test[,1]), col = 2, lty = 1)

# In some examples, pre-fitting models for Y and W separately may
# be helpful (e.g., if different models use different covariates).
# In some applications, one may even want to get Y.hat and W.hat
# using a completely different method (e.g., boosting).

# Generate new data.
n = 4000; p = 20
X = matrix(rnorm(n * p), n, p)
TAU = 1 / (1 + exp(-X[, 3]))
W = rbinom(n ,1, 1 / (1 + exp(-X[, 1] - X[, 2])))
Y = pmax(X[, 2] + X[, 3], 0) + rowMeans(X[, 4:6]) / 2 + W * TAU + rnorm(n)

forest.W = regression_forest(X, W, tune.parameters = TRUE)
W.hat = predict(forest.W)$predictions

forest.Y = regression_forest(X, Y, tune.parameters = TRUE)
Y.hat = predict(forest.Y)$predictions

forest.Y.varimp = variable_importance(forest.Y)

# Note: Forests may have a hard time when trained on very few variables
# (e.g., ncol(X) = 1, 2, or 3). We recommend not being too aggressive
# in selection.
selected.vars = which(forest.Y.varimp / mean(forest.Y.varimp) > 0.2)

tau.forest = causal_forest(X[, selected.vars], Y, W,
                           W.hat = W.hat, Y.hat = Y.hat,
                           tune.parameters = TRUE)

# Check whether causal forest predictions are well calibrated.
test_calibration(tau.forest)

Developing

In addition to providing out-of-the-box forests for quantile regression and causal effect estimation, GRF provides a framework for creating forests tailored to new statistical tasks. If you'd like to develop using GRF, please consult the algorithm reference and development guide.

References

Susan Athey, Julie Tibshirani and Stefan Wager. Generalized Random Forests, Annals of Statistics, forthcoming. [arxiv]

halflearned / grf