What's in here?¹

(1) You'll find material on Machine learning (ML) for problems where one's aim is to measure the effects of some more or less controlled $D$ on some $Y$ in the presence of confounders $X$. ML techniques will be discussed as I learn them, with as objective to show when they can be useful to causal inference. (2) In the different code comments I pay particular attention to differences between ML and Econometrics from the perspective of causal inference. I do not have as objective to compare ML and Econometrics for prediction $E(Y|X)$, but to see how ML methods can be used in Econometrics for estimating causal effects $E(Y(1)-Y(0)|e(X))$, where $e(X)$ may be continuous in $X$; e.g., $e(X)\equiv X$, $X$ a vector of real r.v. (3) You'll also find critical references on how AI is changing our perception of both science methods (a preference for correlation over causation, which in my view is problematic) and the way we live in society (AI deprives humans of part of their labour force, the intellect, namely the ability to generate knowledge from knowledge). Why "critical"? Because AI enthousiasts may have moved too far right recently to be followed blindly.

Regarding (1), I rely on:

• Géron, A., 2022. Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow, O'Reilly, which applies instance- and model-based learning methods to economic and other data.
• Machine learning in Python with scikit-learn, FUN-MOOC.
• Rosenbaum, P., Rubin, D., 1983. The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1), 41-55.

As to point (2), I'll manage to build some kind of dictionary within which one can find words from ML and their translation into econometrics. A few examples: (i) features, factors, explanatory variables are the same thing, (ii) the method of Least Squares used in Econometrics to estimate a model parameters is used for training in ML, (iii) the sum of squared residuals is the cost function, (iv) Overfitting relates to what econometricians call model saturation when specifying and econometric model. A saturated model predicts very well. But saturing a model, though it can be a good control technique, is not key in causal inference. Causal inference requires assumptions and relies on matching techniques where ML can be useful. Some references may help to link the jargon and methods of ML and eocnometrics:

• Athey, S., Imbens, G., 2019. Machine learning methods that economists should know about, Lien.

I firmly believe there are plenty of econometric methods that data scientists should know about. Before to merge the two disciplines, first I need to learn ML methods, which will take a few years.

On (3):

• Le Cun, Y., 2014. Quand la machine apprend, Odile Jacob.
• Mallat, S., 2018. Sciences des données et apprentissage en grande dimension – Leçons inaugurales du Collège de France, fayard.
• Sadin, E., 2023. La vie spectrale – Penser l'ère du métavers et des IA génératives, Grasset.

Programs

• Perceptron for a binary classification.
• Linear regression with a test set (no validation set?).
• Classification without a test set.
• Classification with a test set.
• Classification, Semi-Supervised Learning.

Footnotes

1. Acknowledgments. I am grateful to Alexandre Mutel. ↩