R package: ATE

The main aim of ATE is to provide a user-friendly interface for nonparametric efficient inference of average treatment effects for observational data. The package provides point estimates for average treatment effects, average treatment effect on the treated and can also handle the case of multiple treatments. The package also allows inference by consistent variance estimates.

Requirements

• R (>=3.2.0)
• Rcpp (>=0.12.0)
• RcppArmadillo
• Matrix

Updates

• 2015/09/13 Version 0.4.0 introduced. Source code annotated using Google R style guide.

---

Installation

The package can be installed from CRAN:

install.packages("ATE")

Alternatively, we can directly install from Github using the devtools package:

library(devtools)
install_github("asadharis/ATE")

Key Features

• Ease of use: The main function ATE requires only a numeric matrix X of covariates, numeric vector Y of response and treat vector indicating treatment assignment.

set.seed(1)
library(ATE)
#Generate some data
n <- 500
X1 <- matrix(rnorm(n * 5), ncol = 5)
X2 <- matrix(rbinom(3 * n, 1, 0.4), ncol = 3)
X <- cbind(X1, X2)
prop <- 1 / (1 + exp(X[, 1] - 0.5 * X[, 2] + 0.25 * X[, 3] + X[, 6] + 0.5 * X[, 8]))
treat<- rbinom(n, 1, prop)
Y<-  10 * treat + (2 * treat - 1) *
     (X[, 1] - 0.5 * X[, 2] + 0.25 * X[, 3] + X[, 6] + 0.5 * X[, 8]) + rnorm(n)

#Fit ATE object
fit1 <- ATE(Y, treat, X)
summary(fit1)
Call:
ATE(Y = Y, treat = treat, X = X)

         Estimate Std. Error 95%.Lower 95%.Upper z value   p value    
E[Y(1)] 10.650818   0.112995 10.429353 10.872284 94.2594 < 2.2e-16 ***
E[Y(0)] -0.708631   0.088772 -0.882621 -0.534641 -7.9826 1.433e-15 ***
ATE     11.359449   0.169154 11.027913 11.690986 67.1544 < 2.2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

• plot function for demonstrating effect of covariate balancing for continuous and binary covariates.

plot(fit1)

• We can also estimate the average treatment effect on the treated.

fit2 <- ATE(Y, treat, X, ATT = TRUE)
summary(fit2)
Call:
ATE(Y = Y, treat = treat, X = X, ATT = TRUE)

             Estimate Std. Error 95%.Lower 95%.Upper z value p value    
E[Y(1)|T=1]  9.820802   0.114407  9.596569 10.045035 85.8412  <2e-16 ***
E[Y(0)|T=1]  0.158785   0.127597 -0.091301  0.408870  1.2444  0.2133    
ATT          9.662018   0.214933  9.240757 10.083278 44.9537  <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

• ATE automatically detects and estimates the case of multiple treatment arm.

treat <- rbinom(n, 3, prop)
Y<-  10 * treat + (2 * treat - 1) *
     (X[, 1] - 0.5 * X[, 2] + 0.25 * X[, 3] + X[, 6] + 0.5 * X[, 8]) +
     rnorm(n)
fit3<-ATE(Y,treat,X)
summary(fit3)

Call:
ATE(Y = Y, treat = treat, X = X)

         Estimate Std. Error 95%.Lower 95%.Upper  z value p value    
E[Y(0)] -0.625055   0.114586 -0.849640 -0.400470  -5.4549 4.9e-08 ***
E[Y(1)] 10.559242   0.084657 10.393317 10.725168 124.7291 < 2e-16 ***
E[Y(2)] 22.231546   0.241661 21.757899 22.705194  91.9946 < 2e-16 ***
E[Y(3)] 33.240013   0.352811 32.548516 33.931510  94.2148 < 2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
plot(fit3)

• ATE uses the R packages Rcpp and RcppArmadillo to improve run-time. This allows us to handle big data efficiently. Below we present the example for 10,000 observations and 800 covariates on an Intel® Core™ i5-3337U Processor.

n <- 10000
X1 <- matrix(rnorm(n * 500), ncol = 500)
X2 <- matrix(rbinom(300 * n, 1, 0.4), ncol = 300)
X <- cbind(X1, X2)
prop <- 1 / (1 + exp( X[, 1] - 0.5 * X[, 2] + 0.25 * X[, 3] +  X[, 6] + 0.5 * X[, 8]))
treat<- rbinom(n, 1, prop)
Y<-  10 * treat + (2 * treat - 1) * 
     (X[, 1] - 0.5 * X[, 2] + 0.25 * X[, 3] + X[, 6] + 0.5 * X[, 8]) +
     rnorm(n)

system.time(fit4 <- ATE(Y, treat, X))
   user  system elapsed 
  80.86    2.04   87.55

---

##Installation

• From CRAN: install.packages("ATE") currently version 0.2.0. Slow version without RcppArmadillo.
• From Github: devtools::install_github("asadharis/ATE") latest development version.

---

##Acknowledgements I would like to express my deep gratitude to Professor Gary Chan, my research supervisor, for his patient guidance, enthusiastic encouragement and useful critiques of this project.