College of Education and Behavioral Science, Department of Applied Statistics and Research Methods
University of Northern Colorado
Type of Resources
Place of Publication
University of Northern Colorado
The Sample splitting method in semiparametric statistics could introduce inconsistency in inference and estimation. Thus, to make adaptive learning based on observational data and establish valid learning that helps in the estimation and inference of the parameters and hyperparameters using double machine learning, this study introduces an efficient sample splitting technique for causal inference in the semiparametric framework, in other words, the support points sample splitting( SPSS), a subsampling method based on the energy distance concept is employed for causal inference under double machine learning paradigm. This work is based on the idea that the support points sample splitting (SPSS) is an optimal representative point of the data in a random sample versus the counterpart of random splitting, which implies that the support points sample splitting is an optimal sub-representation of the underlying data generating distribution. To my best knowledge, the conceptual foundation of the support points-based sample splitting is a cutting-edge method of subsampling and the best representation of a full big data set in the sense that the unit structural information of the underlying distribution via the traditional random data splitting is most likely not preserved. Three estimators were applied for double/debiased machine learning causal inference a paradigm that estimates the causal treatment effect from observational data based on machine learning algorithms with the support points sample splitting (SPSS). This study is considering Support Vector Machine (SVM) and Deep Learning (DL) as the predictive estimators. A comparative study is conducted between the SVM and DL with the support points technique to the benchmark results of Chernozhukov et al. (2018) that used instead, the random forest, the neural network, and the regression trees with random k-fold cross-fitting technique. An ensemble machine learning algorithm is proposed that is a hybrid of the super learner and the deep learning with the support points splitting to compare it to the results of Chernozhukov et al. (2018). Finally, a socio-economic real-world dataset, for the 401(k)-pension plan, is used to investigate and evaluate the proposed methods to those in Chernozhukov et al. (2018). The result of this study was under 162 simulations, shows that the three proposed models converge, support vector machine (SVM) with support points sample splitting (SPSS) under double machine learning (DML), the deep learning (DL) with support points sample splitting under double machine learning (DML), and the hybrid of super learning (SL) and deep learning with support points sample splitting under double machine learning. However, the performance of the three models differs. The first model, support vector machine (SVM) with support points sample splitting (SPSS) under double machine learning (DML) has the lowest performance compared to the other two models. In terms of the quality of the causal estimators, it has a higher MSE and inconsistency of the simulation results on all three data dimension levels, low-high-dimensional (p = 20,50,80), moderate-high-dimensional (p = 100, 200, 500), and big-high-dimensional p = (1000, 2000, 5000). The two other models, deep learning (DL) with support points sample splitting under double machine learning (DML), and the hybrid of super learning (SL) and deep learning with support points sample splitting under double machine learning have produced a competing performance and results in terms of the best estimation compared to the two other methods. The first model was time efficient to estimate the causal inference compared to the third one. But the third model was better performing in terms of the estimation quality by producing the lowest MSE compared to the other two models. The results of this research are consistent with the recent development of machine learning. The support vector machine learning has been introduced in the previous century, and it looks like it is no longer showing efficiency and quality estimation with the recent emerging double machine learning. However, cutting-edge methods such as deep learning and super learner have shown superior performance in the estimation of the causal double machine learning target estimator, and efficiency in the time of computation.
Copyright is held by the author.