pattc_deeplearning: Deep PATT-C

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

Deep PATT-C

Description

This function implements the Deep PATT-C method for estimating the Population Average Treatment Effect on the Treated Compliers (PATT-C) using deep learning models using keras and Tensorflow. It consists of training a deep learning model to predict compliance among treated individuals, predicting compliance in the experimental data, training a response model among predicted compliers, and estimating counterfactual outcomes in the population data.

Usage

pattc_deeplearning(
  response.formula,
  compl.var,
  treat.var,
  exp.data,
  pop.data,
  compl.algorithm = "adam",
  response.algorithm = "adam",
  compl.hidden.layer = c(4, 2),
  response.hidden.layer = c(4, 2),
  compl.hidden_activation = "relu",
  response.hidden_activation = "relu",
  response.output_activation = "linear",
  response.output_units = 1,
  response.loss = "mean_squared_error",
  response.metrics = "mean_absolute_error",
  ID = NULL,
  weights = NULL,
  cluster = NULL,
  compl.epoch = 10,
  response.epoch = 10,
  compl.validation_split = NULL,
  response.validation_split = NULL,
  compl.patience = NULL,
  response.patience = NULL,
  compl.dropout_rate = NULL,
  response.dropout_rate = NULL,
  verbose = 1,
  batch_size = 32,
  nboot = 1000,
  seed = 1234
)

Arguments

response.formula	formula specifying the response variable and covariates.
compl.var	string specifying the name of the compliance variable.
treat.var	string specifying the name of the treatment variable.
exp.data	data frame containing the experimental data.
pop.data	data frame containing the population data.
compl.algorithm	string for name of optimizer algorithm for complier model. For optimizers available see keras package.
response.algorithm	string for name of optimizer algorithm for response model. For optimizers available see keras package.
compl.hidden.layer	vector specifying the hidden layers in the complier model and the number of neurons in each hidden layer.
response.hidden.layer	vector specifying the hidden layers in the response model and the number of neurons in each hidden layer.
compl.hidden_activation	string or vector for name of activation function for hidden layers complier model. Defaults to "relu" (Rectified Linear Unit)
response.hidden_activation	string or vector for name of activation function for hidden layers complier model. Defaults to "relu" (Rectified Linear Unit)
response.output_activation	string for name of activation function for output layer of response model. "linear" is recommended for continuous outcome variables, and "sigmoid" for binary outcome variables. For activation functions available see keras package.
response.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.
response.loss	string for loss function in response model. "mean_squared_error" recommended for linear models, "binary_crossentropy" for binary models.
response.metrics	string for metrics in response model. "mean_squared_error" recommended for linear models, "binary_accuracy" for binary models.
ID	optional string specifying the name of the identifier variable.
weights	optional string specifying the name of the weights variable.
cluster	optional string specifying the name of the clustering variable.
compl.epoch	Integer for the number of epochs for complier model.
response.epoch	integer for the number of epochs for response model.
compl.validation_split	double for the proportion of test data to be split as validation in complier model. Defaults to 0.2.
response.validation_split	double for the proportion of test data to be split as validation in response model. Defaults to 0.2.
compl.patience	integer for number of epochs with no improvement after which training will be stopped in complier model.
response.patience	integer for number of epochs with no improvement after which training will be stopped in response model.
compl.dropout_rate	double or vector for proportion of hidden layer to drop out in complier model.
response.dropout_rate	double or vector for proportion of hidden layer to drop out in response model.
verbose	integer specifying the verbosity level during training. Defaults to 1.
batch_size	integer specifying the batch size for training the deep learning models. Default is 32.
nboot	integer specifying the number of bootstrap samples if bootstrap is TRUE. Default is 1000.
seed	random seed

Value

pattc_deeplearning object containing the fitted models, predictions, counterfactuals, and PATT-C estimate.

Examples

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

data("exp_data")
data("pop_data")
set.seed(1243)
deeppattc <- pattc_deeplearning(response.formula = support_war ~ age + female +
income + education +  employed + married + hindu + job_loss,
exp.data = exp_data, pop.data = pop_data,
treat.var = "strong_leader",  compl.var = "compliance",
compl.algorithm = "adam", response.algorithm = "adam",
compl.hidden.layer = c(4,2),  response.hidden.layer = c(4,2),
compl.hidden_activation = "relu",  response.hidden_activation = "relu",
response.output_activation = "sigmoid", response.output_units = 1, 
response.loss = "binary_crossentropy", response.metrics = "accuracy",
compl.epoch = 50, response.epoch = 80,
verbose = 1, batch_size = 32, 
compl.validation_split = 0.2, response.validation_split = 0.2,
compl.dropout_rate = 0.1, response.dropout_rate = 0.1,
compl.patience = 20, response.patience = 20,
nboot = 1000, seed = 1234)

## End(Not run)

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