This repository houses R scripts and data sets that can be used to reproduce the results presented in the paper Accounting for shared covariates in semi-parametric Bayesian additive regression trees. The Annals of Applied Statistics (to appear, 2024).
In addition, it provides an implementation of CSP-BART in the format of an R package named spbart.
library(devtools)
install_github("ebprado/CSP-BART/cspbart",ref='main')library(cspbart)
rm(list = ls())
# ---------------------------------
# CSP-BART for a continuous response
# ---------------------------------
# Simulate from Friedman equation
friedman_data = function(n, num_cov, sd_error){
x = matrix(runif(n*num_cov),n,num_cov)
y = 10*sin(pi*x[,1]*x[,2]) + 20*(x[,3]-.5)^2+10*x[,4]+5*x[,5] + rnorm(n, sd=sd_error)
return(list(y = y,
x = as.data.frame(x)))
}
n = 200
ncov = 5
var = 1
data = friedman_data(n, ncov, sqrt(var))
y = data$y # response variable
X1 = as.data.frame(cbind(y,data$x)) # all covariates + the response
X2 = data$x # all covariates
# Run the semi-parametric BART (WITHOUT intercept)
cspbart.fit = cspbart(formula = y ~ 0 + V4 + V5, x1 = X1, x2 = X2, ntrees = 10, nburn = 2000, npost = 1000)
# Run the semi-parametric BART (WITH intercept)
# cspbart.fit = cspbart(formula = y ~ V4 + V5, x1 = X1, x2 = X2, ntrees = 10, nburn = 2000, npost = 1000)
# Calculate the predicted values (yhat) and parameter estimates (betahat)
yhat = colMeans(spbart.fit$y_hat)
betahat = colMeans(spbart.fit$beta_hat)
# Predict on a new dataset
yhat_pred = predict(cspbart.fit, newdata_x1 = X1, newdata_x2 = X2, type = 'mean')
cor(yhat,yhat_pred) == 1
# Plot
plot(y, yhat);abline(0,1)
plot(1:2, c(10,5), main = 'True versus estimates', ylim=c(3,12))
points(1:2, betahat, col=2, pch=2)
# -----------------------------
# CSP-BART for a binary response
# -----------------------------
n = 200
ncov = 5
var = 1
data = friedman_data(n, ncov, sqrt(var))
aux = data$y
y = ifelse(aux > median(aux), 1, 0)
X1 = as.data.frame(cbind(y,data$x)) # all covariates + the response
X2 = data$x # all covariates
# Run the semi-parametric BART (WITH intercept)
cspbart.fit = cspbart::cl_cspbart(formula = y ~ V4 + V5, x1 = X1, x2 = X2, ntrees = 1, nburn = 2000, npost = 1000)
# Calculate the predicted values (yhat) and parameter estimates (betahat)
yhat = colMeans(pnorm(cspbart.fit$y_hat))
betahat = colMeans(cspbart.fit$beta_hat)
# Predict on a new dataset
yhat_pred = predict(spbart.fit, newdata_x1 = X1, newdata_x2 = X2, type = 'mean')
cor(yhat,yhat_pred) == 1