Causal machine learning: opportunities, limits and social science applications

In the past decade, influental research project have been carried out in two subfields of statistical theory – all of them questioning the basic tenets of classical statistical analysis. On one hand, structural causal modeling presents new ways of doing social research (Morgan & Winship 2015). On the other hand, the application of contemporary machine learning methods has started to revolutionalize econometrics (Athey & Imbens 2019). The two fields seem to indicate contrasting approaches.Structural causal modeling require stronger theoretical assumptions than most of the popular classical statistical techniques. Machine learnig, in contrast, is a data driven modeling approach. However, in practice, those two areas are intertwined. The two fileds have been particularly strongly connected in econometrics in the past five years. Data scientists supporting business decision making are eagerly waiting for causal modeling techniques (Hünermund et al 2021). On the other hand, structural causal modeling itself was born out of the critiques of the data driven approach to AI modeling. Among the leading proponents of causal ML , one can find some of the most influental AI experts of our time (e.g. Bottou 2014, Pearl & Mackenzie 2018, Schölkopf 2019). This causal movement, proposing the development of “machine reasoning” is an important branch of current AI development. Moreover, some of the most cited new results in tha past 3-4 years have been linked to econometric research (Chernozhukov et al 2018, Wager & Athey 2018) and social science (and epidemiology) applications (Chernozhukov et al 2020, Richens et al. 2020). Note however, that there are intensive debates around the new approaches and techniques both among computer scientists and between computer scientists and social scientists. Our research aims at, first, to understand the above methods and the debates around them, and disseminate the new approach in the TK, other MILAB institutions and other social science departments. Our scope is not only to disseminate but also to exchange ideas with data scientists using the data driven aprroaches. We have established a reading group on causal analysis in the fall semester of 2020. We intend to continue the work within this new framework. We would like to collaborate with junior data scientists to combine ideas from computer science and social science statistics.

Keywords: causal machine learning, structural causal models, machine reasoning, business and policy decision making