causalBoosting: Fit a causal boosting model

In saberpowers/causalLearning: Methods for heterogeneous treatment effect estimation

Description

Fit a causal boosting model

Usage

causalBoosting(x, tx, y, num.trees = 500, maxleaves = 4, eps = 0.01,
  splitSpread = 0.1, x.est = NULL, tx.est = NULL, y.est = NULL,
  propensity = FALSE, stratum = NULL, stratum.est = NULL,
  isConstVar = TRUE)

Arguments

x	matrix of covariates
tx	vector of treatment indicators (0 or 1)
y	vector of response values
num.trees	number of shallow causal trees to build
maxleaves	maximum number of leaves per causal tree
eps	learning rate
splitSpread	how far apart should the candidate splits be for the causal trees? (e.g. splitSpread = 0.1) means we consider 10 quantile cutpoints as candidates for making split
x.est	optional matrix of estimation-set covariates used for honest re-estimation (ignored if tx.est = NULL or y.est = NULL)
tx.est	optional vector of estimation-set treatment indicators (ignored if x.est = NULL or y.est = NULL)
y.est	optional vector of estimation-set response values (ignored if x.est = NULL or y.est = NULL)
propensity	logical: should propensity score stratification be used?
stratum	optional vector giving propensity score stratum for each observation (only used if propensity = TRUE)
stratum.est	optional vector giving propensity score stratum for each estimation-set observation (ignored if x.est = NULL or tx.est = NULL or y.est = NULL)
isConstVar	logical: for the causal tree splitting criterion (T-statistc), should it be assumed that the noise variance is the same in treatment and control arms?

Details

This function exists primarily to be called by cv.causalBoosting because the num.trees parameter generally needs to be tuned via cross-validation.

Value

an object of class causalBoosting with attributes:

• CBM: a list storing the intercept, the causal trees and eps

• tauhat: matrix of treatment effects for each patient for each step

• G1: estimated-treatment conditional mean for each patient

• G0: estimated-control conditional mean for each patient

• err.y: training error at each step, in predicting response

• num.trees: number of trees specified by function call

Examples

# Randomized experiment example

n = 100 # number of training-set patients to simulate
p = 10  # number of features for each training-set patient

# Simulate data
x = matrix(rnorm(n * p), nrow = n, ncol = p) # simulate covariate matrix
tx_effect = x[, 1] + (x[, 2] > 0) # simple heterogeneous treatment effect
tx = rbinom(n, size = 1, p = 0.5) # random treatment assignment
y = rowMeans(x) + tx * tx_effect + rnorm(n, sd = 0.001) # simulate response

# Estimate causal boosting model
fit_cb = causalBoosting(x, tx, y, num.trees = 500)
pred_cb = predict(fit_cb, newx = x, num.trees = 500)

# Visualize results
plot(tx_effect, pred_cb, main = 'Causal boosting',
 xlab = 'True treatment effect', ylab = 'Estimated treatment effect')
abline(0, 1, lty = 2)

saberpowers/causalLearning documentation built on May 30, 2019, 8:26 a.m.