MLbalance implements a novel machine learning balance test, the balance permutation test, for experiments with binary, multiarm, and continuous treatments. The purpose of this test is to detect failures of random assignment, data fabrication, or imbalance across treatment arms. For more detail, see Rametta and Fuller (2024).
This package is in development, any recommendations or comments welcome in the issues section.
Stable version is be available on Github here:
# install.packages("pak")
pak::pak("MLCause/MLbalance")You can install the development version of MLbalance from GitHub with:
# install.packages("pak")
pak::pak("CetiAlphaFive/MLbalance")Here is a basic example demonstrating the balance permutation test on a simulated binary treatment DGP with multidimensional contamation of the treatment assignment.
# install.packages("randomizr")
library(MLbalance)
#
set.seed(1995)
#
# Simple simulation
n <- 1000
p <- 20
X <- matrix(rnorm(n*p,0,1),n,p)
w_real <- rbinom(n, 1, ifelse(.021 + abs(.4*X[,4] - .5*X[,8]) < 1, .021 + abs(.4*X[,4] - .5*X[,8]), 1))
# install.packages("randomizr")
w_sim <- randomizr::complete_ra(N = n,m = sum(w_real))
e <- rnorm(n,0,1)
y <- 2*w_real*X[,4] + 3*X[,2] -2*X[,8] + e
df <- data.frame(y,w_real,w_sim,X)
#
r.check <- random_check(W_real = df$w_real, #real treatment assignment vector
W_sim = df$w_sim, #simulated vector, comment out this argument to use permutated real assignment vector instead
X = subset(df,select = -c(y,w_real,w_sim)) #matrix of pretreatment covariates (or any covariates that SHOULD NOT be related to the assignment process/mechanism
); r.check$plot
#> Simulated Assignment Vector Provided, Null Distribution Generated Using Simulated Treatment Assignment.
#>
#>
#> Simple Count Table(s)
#>
#> W_real
#> 0 1
#> 520 480
#> W_sim
#> 0 1
#> 520 480
#>
#>
#> Warning: Extreme Propensity scores detected (greater than .9 or less than .1).
#> Examine $treat.props for more detail.
# to see variables important for predicting assignment, check r.check$imp.predictors