Post Double Selection with Double Machine Learning

Implements post double selection using double machine learning, see Chernozhukov, Chetverikov, Demirer, Duflo, Hansen, Newey, Robins (2017) , for various models, using several back ends. Allows the user flexibility in specifying their own methods for estimation.

postDoubleR

postDoubleR implements post-double selection, also known as double machine learning/debiasing, due Chernozukov et al.(2017a) https://doi.org/10.1111/ectj.12097.

Installation

You can install the stable version of postDoubleR from CRAN with

install.packages("postDoubleR")

The latest version is also available from github using the devtools package

devtools::install_github("JSzitas/postDoubleR")

Introduction

The goal of this package is to facilitate the estimation of treatment effects through double machine learning. To that end, several backends have been implemented in the package, including random forests (through package grf), lasso (through glmnet) and a custom method for declaring any methods to use.

Example

Here we use 5 fold crossfitting, using 100 simulations and lasso (through glmnet) to predict the treatment effects.

library(postDoubleR)

# random data generating process, shamelessly taken from the examples on the grf github page,
# due Tibshirani, Wager, Athey
 n = 2000
 p = 10
 X = matrix(rnorm(n*p), n, p)
 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)


double_ML(X = X,
          Y = Y,
          W = W,
          method = "glmnet",
          lambda.set.Y = 1,
          lambda.set.W = 1,
          k.fld = 5,
          simulations = 100,
          validate.inputs = TRUE,
          seed.use = 1071)

  
 

Here is a simple way to extend the package functionality with custom methods, using the custom_helper function. (glm is an example as this requires no addtional packages.)

# first specify an expression to evaluate for both W and Y
 W_mod <- expression( glm( W~.,
                            family = "binomial",
                            data = as.data.frame(cbind(X,W))))

  Y_mod <- expression( glm( Y~.,
                            family = "gaussian",
                            data = as.data.frame(cbind(X,Y))))
 # then just pass these as custom_helper arguments 
double_ML(X = X,
          Y = Y,
          W = W,
          method = "custom",
          specify.own = custom_helper( X = X,
                                       Y = Y,
                                       W = W,
                                       Y.hat.model = Y_mod,
                                       W.hat.model = W_mod),
          k.fld = 5,
          simulations = 100,
          validate.inputs = TRUE,
          seed.use = 1071)

Note that you can compose more complicated functions, and they should work if they are loaded in your environment.

  composed_Y <- function(X,Y){
    X <- as.data.frame(X)
    return(glm( Y~.,
                family = "gaussian",
                data = as.data.frame(cbind(X,Y))))
  }

  composed_W <- function(X,Y){
    X <- as.data.frame(X)
    return(glm( Y~.,
                family = "gaussian",
                data = as.data.frame(cbind(X,Y))))
  }

  W_mod <- expression( composed_W(X,W) )

  Y_mod <- expression( composed_Y(X,Y) )

  
double_ML(X = X,
          Y = Y,
          W = W,
          method = "custom",
          specify.own = custom_helper( X = X,
                                       Y = Y,
                                       W = W,
                                       Y.hat.model = Y_mod,
                                       W.hat.model = W_mod),
          k.fld = 5,
          simulations = 100,
          validate.inputs = TRUE,
          seed.use = 1071)
  

Contributing

If you would like to contribute a pull request, please do contribute! All contributions will be considered for acceptance, provided they are justifiable and the code is reasonable, regardless of anything related to the person submitting the pull request. Please keep things civil - there is no need for negativity. Also, please do refrain from adding unnecessary dependencies (Ex: pipe) to the package (such pull requests as would add an unnecessary dependency will be denied/ suspended until the code can be made dependency free).