因果推論下的機器學習
Machine Learning and Causal Inference
國立政治大學經濟學系(Department of Economics, National Chengchi University)
摘要
傳統機器學習方法著重於提升預測準確度,但近年研究顯示,這些方法亦可應用於高維度資料中,協助選取控制變數,進而提高因果推論的可信度。本文首先說明預測與因果推論在研究目標與方法設計上的根本差異,接著介紹兩種整合方法:Post-Double Selection LASSO 與Double/Debiased Machine Learning。這兩種方法皆透過機器學習,在高維度資料下輔助因果推論。本文亦簡要介紹兩個實證上的應用,分別探討生育補助對生育率的影響,以及制度對經濟成長的影響,以說明機器學習在因果推論中的實際運用。
Abstract
Traditional machine learning methods focus on improving predictive accuracy. Recent research, however, shows that these methods can also be applied to high-dimensional data to assist in selecting control variables, thereby improving the credibility of causal inference. This article first explains the fundamental differences between prediction and causal inference in terms of research objectives and methodology. It then introduces two integrated approaches: Post-Double Selection LASSO and Double/Debiased Machine Learning. Both methods use machine learning to support causal inference in high-dimensional settings. The article also briefly discusses two empirical applications: the effect of fertility subsidies on fertility rates and the effect of institutions on economic growth. These examples illustrate the practical use of machine learning in causal inference.
關鍵詞
- 類型
- 論文
- 期刊
- 經濟論文叢刊
- ISSN
- 1018-3833
- 卷期
- 54卷2期
- 出版
- 2026年06月
- 頁碼
- 171–209
- 編號
- NTUEJ00006563
同期目錄
- 因果推論下的機器學習黃柏鈞(Po-Chun Huang)・頁 171–209
- 就讀台北市明星高中對進入好大學的影響劉家樺(Chia-Hua Liu);駱明慶(Ming-Ching Luoh)・頁 211–266
- 台灣勞動市場配對陳冠銘(Kuan-Ming Chen);蔡霖東(Lin-Tung Tsai);楊昊紘(Hau-Hung Yang)・頁 267–289