The xtdml package implements double machine learning
(DML) for static partially linear regression models for panel data with
fixed effects, as in Clarke
and Polselli (2025). The xtdml package is built on the
DoubleML package by Bach et al. (2022) using the
mlr3 ecosystem and the same notation.
The xtdml package estimates the structural (causal)
parameter from panel data models:
Y_{it} = \theta_0 D_{it} + g_0(X_{it}) + \alpha_i + U_{it} D_{it} = m_0(X_{it}) + \gamma_i + V_{it},where * \(Y_{it}\) is the output, \(D_{it}\) the treatment, \(X_{it}\) the covariates * \(\theta_0\) is the structural (causal) parameter to estimate; * \((l_0, m_0)\) are (possibly nonlinear) nuisance functions to learn from the data using one of the tree proposed approaches (“fd-exact”, “wg-approx”, “cre”); * (\(\alpha_i, \gamma_i\)) are the unobserved individual heterogeneity correlated with the included covariates; * (\(U_{it}\), \(V_{it}\)) are disturbances.
[!NOTE] The current version of the package allows the user to immediately use the estimation tools without the need to proceed with additional data managing and transformations. In particular, the user can choose:
The panel data approach to use among
approach = ("fd-exact", "wg-approx", "cre"); default is"fd-exact".xtdmlproceeds with transforming the data based on the selected approach, following Clarke and Polselli (2025).The type of transformation to apply to the covariates \(X\) in the data set among
transformX = ("no", "minmax", "poly")."no"does not transform the covariatesXand is recommended for tree-based learners."minmax"applies the Min-Max normalization \(x' = (x-x_{min})/(x_{max}-x_{min})\) to the covariates and is recommended with neural networks."poly"add polynomials up to order three and interactions between all possible combinations of two and three variables; this is recommended for Lasso. Default is"no".
[!CAUTION] 1. The package has been renamed
xtdmlfromXTDML! <<<<<<< HEAD 2. Major changes in xtdml_data_from_data_frame() function have been introduced since version 0.1.7 and various bugs corrected. 3.tune_settingshave been updated aftermlr3tuning(>= 0.20.0),mlr3tuning(>= 0.19.0),bbotk(>= 1.6.0) 4. Previous versions of the package allowed for treatment endogeneity in the partially linear panel regression model. This option has been temporally removed, but we are currently working on including IV estimation and weak IV tests in the current version of thextdmlpackage. (Panel IV DML coming soon) 5. With the current version of the package it is no longer possible to choose the hybrid approach; this is consistent with the published version of the article. =======
Previous versions of the package allowed for treatment endogeneity in the partially linear panel regression model. This option has been temporally removed, but we are currently working on including IV estimation and weak IV tests in the current version of the
xtdmlpackage. (Panel IV DML coming soon)With the current version of the package it is no longer possible to choose the hybrid approach; this is consistent with the published version of the article.
Vignettes will be added soon.
52433c7 (Version 0.1.12)
The installation of xtdml can be done from CRAN
using
install.packages("xtdml")
library(xtdml)or via devtools from GitHub for the latest version
library(devtools)
install_github("POLSEAN/xtdml")<<<<<<< HEAD
library(xtdml)
library(mlr3)
library(mlr3learners)
library(paradox)
library(mlr3tuning)
library(mlr3misc)
library(xgboost)
library(tibble)
library(dplyr)
library(datawizard)
library(data.table)
# Suppress warning messages
lgr::get_logger("mlr3")$set_threshold("warn")
lgr::get_logger("bbotk")$set_threshold("warn")
#______________________________________________________________________________#
# Generate data
#______________________________________________________________________________#
set.seed(1234)
# Load data
df = make_plpr_data(n_obs = 1000, t_per = 10, dim_x = 20, theta = 0.5, rho=0.8)
approaches = c("fd-exact","cre", "wg-approx")
for (approach in approaches){
# Initialize data environment
x_cols = paste0("X", 1:20)
obj_xtdml_data = xtdml_data_from_data_frame(df,
x_cols = x_cols, y_col = "y", d_cols = "d",
panel_id = "id",
time_id = "time",
approach = approach,
transformX = "no")
print(obj_xtdml_data)
# Gradient boosting
learner = lrn("regr.xgboost", nrounds = 100)
ml_m = learner$clone()
ml_l = learner$clone()
type_score = c("orth-PO", "orth-IV")
print(paste0("approach:", approach))
for (score in type_score){
if(score == "orth-PO"){
xtdml_obj = xtdml_plr$new(obj_xtdml_data,
ml_l = ml_l, ml_m = ml_m,
n_folds = 5,
score = score)
param_grid = list("ml_l" = ps(max_depth = p_int(2,10),
lambda = p_dbl(0,2)),
"ml_m" = ps(max_depth = p_int(2,10),
lambda = p_dbl(0,2)))
}else if (score == "orth-IV"){
ml_g = learner$clone()
xtdml_obj = xtdml_plr$new(obj_xtdml_data,
ml_l = ml_l, ml_m = ml_m, ml_g = ml_g,
n_folds = 5,
score = score)
param_grid = list("ml_l" = ps(max_depth = p_int(2,10),
lambda = p_dbl(0,2)),
"ml_m" = ps(max_depth = p_int(2,10),
lambda = p_dbl(0,2)),
"ml_g" = ps(max_depth = p_int(2,10),
lambda = p_dbl(0,2)))
}
tune_settings = list(n_folds_tune = 5,
rsmp_tune = mlr3::rsmp("cv", folds = 5),
terminator = mlr3tuning::trm("evals", n_evals = 5),
tuner = tnr("grid_search", resolution = 10))
xtdml_obj$tune(param_set = param_grid, tune_settings = tune_settings)
# Fit DML model and print results
xtdml_obj$fit()
#xtdml_obj$print()
xtdml_obj$summary()
ci = xtdml_obj$confint()
}
print(xtdml_obj$params)
table = matrix(0, 10, 1)
table[,1] = cbind(xtdml_obj$coef_theta,
xtdml_obj$se_theta,
xtdml_obj$pval_theta,
ci[ , 1], ci[ , 2],
xtdml_obj$model_rmse,
as.numeric(xtdml_obj$rmses["ml_l"]),
as.numeric(xtdml_obj$rmses["ml_m"]),
xtdml_obj$data$n_obs,
length(unique(xtdml_obj$data$data_model[[xtdml_obj$data$cluster_cols]])))
rownames(table)= c("Estimate", "Std. Error", "P-value", "Lower bound 95% CI", "Upper bound 95% CI",
"Model RMSE", "MSE of l", "MSE of m","No. observations","No. groups")
colnames(table)= c("xtdml-boosting")
formatted_table = matrix("", nrow=10, ncol=1)
formatted_table[1:8, 1] = formatC(table[1:8, 1], digits=4, format="f")
formatted_table[9:10, 1] = formatC(table[9:10, 1], digits=0, format="f")
# Keep row and column names
rownames(formatted_table) = rownames(table)
colnames(formatted_table) = colnames(table)
# Print
print(formatted_table, quote = FALSE)
}
======= >>>>>>> 52433c7 (Version 0.1.12) ## References Bach, P., Chernozhukov, V., Kurz, M. S., Spindler, M. and Klaassen, S. (2024), DoubleML - An Object-Oriented Implementation of Double Machine Learning in R, Journal of Statistical Software, 108(3): 1-56, doi:10.18637/jss.v108.i03.
Bach P, Kurz MS, Chernozhukov V, Spindler M, Klaassen S (2024b). DoubleML: Double Machine Learning in R. R package version 1.0.2, URL https://cran.r-project.org/ web/packages/DoubleML/index.html.
Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 21(1):C1–C68.
Clarke, P. S. and Polselli, A. (2025). Double Machine Learning for Static Panel Models with Fixed Effects. The Econometrics Journal. DOI: 10.1093/ectj/utaf011.
<<<<<<< HEAD Polselli A (2025). xtdml: Double Machine Learning for Static Panel Models with Fixed Effects. R package version 0.1.5, URL https://cran.r-project.org/web/packages/xtdml/ index.html ======= Polselli A (2025). xtdml: Double Machine Learning for Static Panel Models with Fixed Ef- fects. R package version 0.1.5, URL https://cran.r-project.org/web/packages/xtdml/ index.html >>>>>>> 52433c7 (Version 0.1.12)