metalearner_deeplearning: metalearner_deeplearning

In DeepLearningCausal: Causal Inference with Super Learner and Deep Neural Networks

metalearner_deeplearning

Description

metalearner_deeplearning implements the meta learners for estimating CATEs using Deep Neural Networks through Tensorflow. Deep Learning Estimation of CATEs from four meta-learner models (S,T,X and R-learner) using TensorFlow and Keras3

Usage

metalearner_deeplearning(
  data = NULL,
  train.data = NULL,
  test.data = NULL,
  cov.formula,
  treat.var,
  meta.learner.type,
  nfolds = 5,
  algorithm = "adam",
  hidden.layer = c(2, 2),
  hidden_activation = "relu",
  output_activation = "linear",
  output_units = 1,
  loss = "mean_squared_error",
  metrics = "mean_squared_error",
  epoch = 10,
  verbose = 1,
  batch_size = 32,
  validation_split = NULL,
  patience = NULL,
  dropout_rate = NULL,
  conformal = FALSE,
  alpha = 0.1,
  calib_frac = 0.5,
  prob_bound = TRUE,
  seed = 1234
)

Arguments

data	data.frame object of data. If a single dataset is specified, then the model will use cross-validation to train the meta-learners and estimate CATEs. Users can also specify the arguments (defined below) to separately train meta-learners on their training data and estimate CATEs with their test data.
train.data	data.frame object of training data for Train/Test mode. Argument must be specified to separately train the meta-learners on the training data.
test.data	data.frame object of test data for Train/Test mode. Argument must be specified to estimate CATEs on the test data.
cov.formula	formula description of the model y ~ x(list of covariates). Permits users to specify covariates in the meta-learner model of interest. This includes the outcome variables and the confounders.
treat.var	string for name of Treatment variable. Users can specify the treatment variable in their data by employing the treat.var argument.
meta.learner.type	string of "S.Learner", "T.Learner", "X.Learner", or "R.Learner". Employed to specify any of the following four meta-learner models for estimation via deep learning: S,T,X or R-Learner.
nfolds	integer for number of folds for Meta Learners. When a single dataset is specified, then users employ cross-validation to train the meta-learners and estimate CATEs. For a single dataset, users specify nfolds to define the number of folds to split data for cross-validation.
algorithm	string for optimization algorithm. For optimizers available see keras package. Arguments to reconfigure and train the deep neural networks for meta-learner estimation include the optimization algorithm. Options for the optimization alogrithm include "adam", “adagrad”, “rmsprop”, “sgd”.
hidden.layer	permits users to specify the number of hidden layers in the model and the number of neurons in each hidden layer.
hidden_activation	string or vector for name of activation function for hidden layers of model. Defaults to "relu" which means that users can specify a single value to use one activation function for each hidden layer. While "relu" is a popular choice for hidden layers, users can also use "softmax" which converts a vector of values into a probability distribution and "tanh" that maps input to a value between -1 and 1.
output_activation	string for name of activation function for output layer of model. "linear" is recommended for continuous outcome variables, and "sigmoid" for binary outcome variables. For activation functions available see keras package. 'For instance, Keras provides various activation functions that can be used in neural network layers to introduce non-linearity
output_units	integer for units in output layer. Defaults to 1 for continuous and binary outcome variables. In case of multinomial outcome variable, set to the number of categories.
loss	string for loss function "mean_squared_error" recommended for linear models, "binary_crossentropy" for binary models.
metrics	string for metrics in response model. "mean_squared_error" recommended for linear models, "binary_accuracy" for binary models.
epoch	interger for number of epochs. epoch denotes one complete pass through the entire training dataset. Model processes each training example once during an epoch.
verbose	integer specifying the verbosity level during training. 1 for full information and learning curve plots. 0 to suppress messages and plots.
batch_size	integer for batch size to split training data. batch size refers to the number of training samples processed before the model's parameters are updated. Batch size is a vital hyperparameter that affects both training speed and model performance. It is crucial for computational efficiency.
validation_split	double for proportion of training data to split for validation. validation split involves partitioning data into training and validation sets to build and tune model.
patience	integer for number of epochs with no improvement to wait before stopping training. patience stops training of neural network if model's performance on validation data stops improving.
dropout_rate	double or vector for proportion of hidden layer to drop out. dropout rate is hyperparameter for preventing a model from overfitting the training data.
conformal	logical for whether to compute conformal prediction intervals conformal prediction intervals provide measure of uncertainty for ITEs.
alpha	proportion for conformal prediction intervals alpha proportion refers to significance level that guarantees desired coverage probability for ITEs
calib_frac	fraction of training data to use for calibration in conformal inference
prob_bound	logical for whether to bound conformal intervals within [-1,1] for classification models
seed	random seed

Value

metalearner_deeplearning object with CATEs

Examples

## Not run: 
#check for python and required modules
python_ready()
data("exp_data")

s_deeplearning <- metalearner_deeplearning(data = exp_data,
cov.formula  = support_war ~ age + female + income + education
+ employed + married + hindu + job_loss, 
treat.var = "strong_leader",  meta.learner.type = "S.Learner",
nfolds = 5,  algorithm = "adam",
hidden.layer = c(2,2),  hidden_activation = "relu",
output_activation = "sigmoid", output_units = 1,
loss = "binary_crossentropy",  metrics = "accuracy",
epoch = 10,  verbose = 1,   batch_size = 32, 
validation_split = NULL,  patience = NULL,
dropout_rate = NULL, conformal= FALSE,  seed=1234)

## End(Not run)
## Not run: 
#check for python and required modules
python_ready()
data("exp_data")

t_deeplearning <- metalearner_deeplearning(data = exp_data,
cov.formula  = support_war ~ age + female + income + education
+ employed + married + hindu + job_loss, 
treat.var = "strong_leader",  meta.learner.type = "T.Learner",
nfolds = 5,  algorithm = "adam",
hidden.layer = c(2,2),  hidden_activation = "relu",
output_activation = "sigmoid", output_units = 1,
loss = "binary_crossentropy",  metrics = "accuracy",
epoch = 10,  verbose = 1,   batch_size = 32, 
validation_split = NULL,  patience = NULL,
dropout_rate = NULL, conformal= TRUE, 
alpha = 0.1,calib_frac = 0.5, prob_bound = TRUE, seed = 1234)

## End(Not run)

DeepLearningCausal documentation built on Nov. 6, 2025, 5:08 p.m.