causalegm {RcausalEGM}

Main function for estimating causal effect in either binary or continuous treatment settings.

Description

This function takes observation data (x,y,v) as input, and estimate the ATE/ITE/ADRF.

Usage

causalegm(
  x,
  y,
  v,
  z_dims = c(3, 3, 6, 6),
  output_dir = ".",
  dataset = "myData",
  lr = 2e-04,
  bs = 32,
  alpha = 1,
  beta = 1,
  gamma = 10,
  g_d_freq = 5,
  g_units = c(64, 64, 64, 64, 64),
  e_units = c(64, 64, 64, 64, 64),
  f_units = c(64, 32, 8),
  h_units = c(64, 32, 8),
  dv_units = c(64, 32, 8),
  dz_units = c(64, 32, 8),
  save_model = FALSE,
  save_res = FALSE,
  binary_treatment = TRUE,
  use_z_rec = TRUE,
  use_v_gan = TRUE,
  random_seed = 123,
  n_iter = 30000,
  normalize = FALSE,
  x_min = NULL,
  x_max = NULL
)

Arguments

x

is the treatment variable, one-dimensional array with size n.

y

is the potential outcome, one-dimensional array with size n.

v

is the covariates, two-dimensional array with size n by p.

z_dims

is the latent dimensions for z_0,z_1,z_2,z_3 respectively. Total dimension should be much smaller than the dimension of covariates v. Default: c(3,3,6,6)

output_dir

is the folder to save the results including model hyperparameters and the estimated causal effect. Default is ".".

dataset

is the name for the input data. Default: "myData".

lr

is the learning rate. Default: 0.0002.

bs

is the batch size. Default: 32.

alpha

is the coefficient for the reconstruction loss. Default: 1.

beta

is the coefficient for the MSE loss of x and y. Default: 1.

gamma

is the coefficient for the gradient penalty loss. Default: 10.

g_d_freq

is the iteration frequency between training generator and discriminator in the Roundtrip framework. Default: 5.

g_units

is the list of hidden nodes in the generator/decoder network. Default: c(64,64,64,64,64).

e_units

is the list of hidden nodes in the encoder network. Default: c(64,64,64,64,64).

f_units

is the list of hidden nodes in the f network for predicting y. Default: c(64,32,8).

h_units

is the list of hidden nodes in the h network for predicting x. Default: c(64,32,8).

dv_units

is the list of hidden nodes in the discriminator for distribution match v. Default: c(64,32,8).

dz_units

is the list of hidden nodes in the discriminator for distribution match z. Default: c(64,32,8).

save_model

whether to save the trained model. Default: FALSE.

save_res

whether to save the results during training. Default: FALSE.

binary_treatment

whether the treatment is binary or continuous. Default: TRUE.

use_z_rec

whether to use the reconstruction loss for z. Default: TRUE.

use_v_gan

whether to use the GAN training for v. Default: TRUE.

random_seed

is the random seed to fix randomness. Default: 123.

n_iter

is the training iterations. Default: 30000.

normalize

whether apply normalization to covariates. Default: FALSE.

x_min

ADRF start value. Default: NULL

x_max

ADRF end value. Default: NULL

Value

causalegm returns an object of class "causalegm".

An object of class "causalegm" is a list containing the following:

causal_pre

the predicted causal effects, which are individual causal effects (ITEs) in binary treatment settings and dose-response values in continous treatment settings.

getCATE

the method for getting the conditional average treatment effect (CATE).It takes covariates v as input.

predict

the method for outcome function. It takes treatment x and covariates v as inputs.

References

Qiao Liu, Zhongren Chen, Wing Hung Wong. CausalEGM: a general causal inference framework by encoding generative modeling. arXiv preprint arXiv:2212.05925, 2022.

Examples

#Generate a simple simulation data.
n <- 1000
p <- 10
v <- matrix(rnorm(n * p), n, p)
x <- rbinom(n, 1, 0.4 + 0.2 * (v[, 1] > 0))
y <- pmax(v[, 1], 0) * x + v[, 2] + pmin(v[, 3], 0) + rnorm(n)
model <- causalegm(x=x, y=y, v=v, n_iter=3000)
paste("The average treatment effect (ATE):", round(model$ATE, 2))