README.md

In siqixu/DeepMediation: DeepMediation: De-biased Deep Learning for Semi-parametric Causal Mediation Analysis

DeepMediation: De-biased Deep Learning for Semi-parametric Causal Mediation Analysis

DeepMediation is an approach for semi-parametric causal mediation analysis to estimate the natrual (in)direct effects of a binary treatment on an outcome of interet. DeepMediation adopts the deep neural networks to estimate the nuisance parameters involved in the influence functions of the potential outcomes.

Setup

The DeepMediation package will use the R package "keras" to establish the neural networks. It also depends on the R package "foreach" for parallel computation. Therefore, make sure that these packages have been installed.

Use the following command in R to install the package:

install.packages(pkgs="keras")  # install the "keras" package
install.packages(pkgs="foreach")  # install the "foreach" package
library(devtools)
install_github("siqixu/DeepMediation",ref="main") # install the "DeepMediation" package

Usage

DeepMed(y,d,m,x,hyper_grid,epochs=500,hyper,cv=TRUE,trim=0.05)

y: The outcome variable in causal mediation analysis.

d: The exposure variable in causal mediation analysis.

m: The mediator variable in causal mediation analysis.

x: The covariates in causal mediation analysis.

hyper_grid: A grid of candidate hyperparameters for deep neural networks. It is a numeric matrix with three columns for numbers of hidden units, numbers of hidden layers, and L1 regularizition parameters in the input layer, respecively. "hyper_grid" should be provided if cv==TRUE, then a 3-fold cross-validation will be used to select the hyperparameters.

epochs: The maximum number of candidate epochs.

hyper: The hyperparameters of neural networks, i.e., L1 regularizition parameters, numbers of hidden layers, numbers of hidden units and epochs. The neural networks with hyperparameters specified by "hyper" will be used in a 3-fold cross-fitting to estiamte the nuisance parameters in the influence functions for the mean potential outcomes. "hyper" should be provided if cv==FALSE.

cv: If cv==TRUE, then "hyper_grid" should be provided and a 3-fold cross-validation will be used to select hyperparameters among "hyper_grid". Otherwise, hyperparameters should be provided in the argument "hyper".

trim: The trimming rate for preventing conditional treatment or mediator probabilities from being zero. Observations with any denominators in the potential outcomes smaller than the trimming rate will be excluded from the analysis.

Value

results: The estimates (effect), standard errors (se) and P values (p-val) of the total treatment effect (total), (in)direct treatment effect in treated ((in)dir.treat), and (in)direct treatment effect in control group ((in)indir.control).

ntrimmed: The number of observations being excluded due to the denominators in the potential outcomes smaller than the trimming rate.

Details

All binary variables in the data should be coded as 0 or 1.

References

Xu S, Liu L and Liu Z. De-biased Deep Learning for Semi-parametric Causal Mediation Analysis. 2021

Acknowledgement

Some functions in "DeepMediation" package are built upon the framework of the "medDML" function from the R package "causalweight". We thank the authors, Hugo Bodory and Martin Huber, of the "causalweight" package.

Examples

library(DeepMediation)
# use parallel computation with 30 cores for a grid search of hyperparameters
library(doMC); registerDoMC(30)

l1 = c(0,0.05,0.1)
layer = c(1:3)
unit = c(10,20,50)
hyper_grid=expand.grid(unit,layer,l1) # create a grid of candidate hyperparameters
DeepMed(y,d,m,x,hyper_grid,cv=TRUE)

siqixu/DeepMediation documentation built on Dec. 23, 2021, 2:30 a.m.