Double Machine Learning Simplified: Part 1 — Basic Causal Inference Applications | by Jacob Pieniazek | Jul, 2023

All Images By Author

Learn how to utilize DML in causal inference tasks

This article is the 1st in a 2 part series on simplifying and democratizing Double Machine Learning. In the 1st part, we will be covering the fundamentals of Double Machine Learning, along with two basic causal inference applications. Then, in pt. 2, we will extend this knowledge to turn our causal inference problem into a prediction task, wherein we predict individual level treatment effects to aid in decision making and data-driven targeting.

The conceptual & practical distinctions between statistical/machine learning (ML) and causal inference/econometric (CI) tasks has been established for years— ML seeks to predict, whereas CI seeks to infer a treatment effect or a “causal” relationship between variables. However, it was common, and still is to a certain degree, for the data scientist to draw causal conclusions from parameters of a trained machine learning model, or some other interpretable ML methodology. Despite this, there has been significant strides in industry and across many academic disciplines to push more rigorousness in making causal claims, and this has stimulated a much wider and open discourse on CI. In this stride, we have seen amazing work come out that has begun to bridge the conceptual gap between ML and CI, specifically tools in CI that take advantage of the power of ML methodologies.

The primary motivation for this series is to democratize the usage of & applications of Double Machine Learning (DML), first introduced by Chernozhukov et al. in their pioneering paper “Double Machine Learning for Treatment and Causal Parameters”, and to enable the data scientist to utilize DML in their daily causal inference tasks.[1] In doing so, we will first dive into the fundamentals of DML. Specifically, we will cover some of the conceptual/theoretical underpinnings, including the regression framework for causality & the Frisch-Waugh-Lovell Theorem, and then we will use this framework to develop DML. Lastly, we will demonstrate two notable applications of Double Machine Learning:

  1. Converging towards Exogeneity/CIA/Ignorability in our Treatment given Non-Experimental/Observational Data, and

Source link

Leave a Comment