Pathwise Estimation of Covariate Balancing Propensity Scores

🚧 Work in progress 🚧

A package for pathwise estimation of regularized logistic propensity score models using covariate balancing loss functions rather than maximum likelihood. Regularization paths are fit via the adelie elastic-net solver with a glmnet-like interface and objectives that directly target covariate balance for the ATE and ATT.

Some helpful links for getting started:

• An introduction to balnet
• Package docs

Installation

The development version can be installed via

devtools::install_github("erikcs/balnet", subdir = "r-package/balnet")

Installing from source requires a C++17 compiler or later. To build with multithreading enabled, OpenMP needs to be available (on Mac, a simple option is to set the default C++ compiler to gcc installed via brew).

Usage Example

# Simulate data with confounding.
n <- 2000
p <- 10
X <- matrix(rnorm(n * p), n, p)
W <- rbinom(n, 1, 1 / (1.5 + exp(X[, 2] + X[, 3])))
Y <- W + 2 * log(1 + exp(X[, 1] + X[, 2] + X[, 3])) + rnorm(n)

# Fit model targeting the ATE = E[Y(1)] - E[Y(0)].
# Two logistic models are fit: one for treated, one for control.
fit <- balnet(X, W, target = "ATE")

# Print path summary.
print(fit)

# Visualize the path.
plot(fit)

# Plot the covariate imbalance at given lambda.
# Note: lambda = 0 selects the final lambda in the sequence. Scalar values
# are applied to both arms.
plot(fit, lambda = 0)

# Predict propensity scores at end of lambda path.
W.hat <- predict(fit, X, lambda = 0)

# Get weights at end of lambda path.
ipw.weights <- ipw(fit, lambda = 0)

# Estimate ATE using IPW weights.
mean(Y * (ipw.weights$treated - ipw.weights$control))