Computational Benefits of Short-Stacking

Introduction

This article illustrates the computational advantages of short-stacking over conventional stacking for estimation of structural parameters using double/debiased machine learning. See also Ahrens et al. (2023, 2024) for further discussion of short-stacking.

Estimation with Stacking and Short-Stacking

We apply ddml to the included random subsample of 5,000 observations from the data of Angrist & Evans (1998). The data contains information on the labor supply of mothers, their children, as well as demographic data. See ?AE98 for details.

# Load ddml and set seed
library(ddml)
set.seed(221945)

# Construct variables from the included Angrist & Evans (1998) data
y = AE98[, "worked"]
D = AE98[, "morekids"]
Z = AE98[, "samesex"]
X = AE98[, c("age","agefst","black","hisp","othrace","educ")]

For a comparison of run-times, we consider the following three estimators for the nuisance parameters arising in estimation of the local average treatment effect (LATE):

1. Gradient boosting as the single base learner (see ?mdl_xgboost)
2. Short-stacking with linear regression (see ?ols), lasso (see ?mdl_glmnet), and gradient boosting
3. Stacking with linear regression, lasso, and gradient boosting

time_singlelearner <- system.time({
  late_fit <- ddml_late(y, D, Z, X,
                        learners = list(what = mdl_xgboost,
                                        args = list(nrounds = 100,
                                                    max_depth = 1)),
                        sample_folds = 10,
                        silent = TRUE)
  })#SYSTEM.TIME

time_shortstacking <- system.time({
  late_fit <- ddml_late(y, D, Z, X,
                        learners = list(list(fun = ols),
                                        list(fun = mdl_glmnet),
                                        list(fun = mdl_xgboost,
                                             args = list(nrounds = 100,
                                                         max_depth = 1))),
                        ensemble_type = 'nnls1',
                        shortstack = TRUE,
                        sample_folds = 10,
                        silent = TRUE)
  })#SYSTEM.TIME

time_stacking <- system.time({
  late_fit <- ddml_late(y, D, Z, X,
                        learners = list(list(fun = ols),
                                        list(fun = mdl_glmnet),
                                        list(fun = mdl_xgboost,
                                             args = list(nrounds = 100,
                                                         max_depth = 1))),
                        ensemble_type = 'nnls1',
                        shortstack = FALSE,
                        sample_folds = 10,
                        cv_folds = 10,
                        silent = TRUE)
  })#SYSTEM.TIME

Both stacking and short-stacking construct weighted averages of the considered base learners to minimize the out-of-sample mean squared prediction error (MSPE). The difference between the two approaches lies in the construction of the MSPE: While stacking runs cross-validation in each cross-fitting sample fold, short-stacking directly uses the out-of-sample predictions arising in the cross-fitting step of double/debiased machine learning estimators. As the run-times below show, this results in a substantially reduced computational burden:

cat("Time single learner:", time_singlelearner[1], "\n")
#> Time single learner: 8.737
cat("Time short-stacking:", time_shortstacking[1], "\n")
#> Time short-stacking: 21.295
cat("Time stacking:      ", time_stacking[1])
#> Time stacking:       124.636

References

Ahrens A, Hansen C B, Schaffer M E, Wiemann T (2023). “ddml: Double/debiased machine learning in Stata.” https://arxiv.org/abs/2301.09397

Ahrens A, Hansen C B, Schaffer M E, Wiemann T (2024). “Model averaging and double machine learning.” https://arxiv.org/abs/2401.01645

Angrist J, Evans W (1998). “Children and Their Parents’ Labor Supply: Evidence from Exogenous Variation in Family Size.” American Economic Review, 88(3), 450-477.