#causalml — Public Fediverse posts

Линейная регрессия на стероидах: Double Machine Learning для устранения смещений в данных

Любой аналитик знает, что самым надёжным способом проверки гипотез являются рандомизированные контролируемые эксперименты (RCT), или, как их называют в народе — A/B-тесты. На практике часто возникают ситуации, когда провести A/B-тест невозможно — в основном это происходит по этическим или техническим причинам. Однако бывают кейсы, когда рандомизация невозможна потому, что treatment-ом является определённое действие пользователя. Например, treatment-ом может быть оформление платной подписки или отмена бронирования на сервисе. Давайте назовём такой вид воздействия добровольным. В русскоязычном пространстве, и в частности на Хабре, достаточно много статей, посвящённых таким методам Causal Inference, как DiD, PSM и Causal Impact. Тем не менее, к моему удивлению, практически нет статей, посвящённых методам на основе ортогонализации и regression adjustment, хотя, на мой взгляд, именно эти методы являются самыми удобными для оценки эффекта от добровольного treatment-а. Пришло время исправить это недоразумение и разобрать метод Double/Debiased Machine Learning (DML) и Partial Linear Regression для задач Causal Inference!

https://habr.com/ru/articles/1043704/

#causal_inference #machine_learning #abтестирование #причинноследственный_анализ #differenceindifference #psm #causalml #causalimpact #causal_effect #causality

I just made my first PR on an #OpenSource project. Ever. Why did that take so long? IDK - I guess all the projects I worked with had everything I needed, and I was too busy to volunteer my time for something I didn't need. The PR is just a small suggestion to improve memory performance in #EconML (#causalml) - we'll see what the repo maintainers think of my hackery, haha...

I just made my first PR on an #OpenSource project. Ever. Why did that take so long? IDK - I guess all the projects I worked with had everything I needed, and I was too busy to volunteer my time for something I didn't need. The PR is just a small suggestion to improve memory performance in #EconML (#causalml) - we'll see what the repo maintainers think of my hackery, haha...

#CausalML update - I am now fitting my first #CausalForest on real data!

Does anyone have advice on the most important #hyperparameters (After the # of trees & tree depth.)

I'm working on large imbalanced data sets and a large number of treatment variables, so it's not like anything you see in the economics literature. 🤔 #ML #AI #causal

#CausalML update - I am now fitting my first #CausalForest on real data!

Does anyone have advice on the most important #hyperparameters (After the # of trees & tree depth.)

I'm working on large imbalanced data sets and a large number of treatment variables, so it's not like anything you see in the economics literature. 🤔 #ML #AI #causal

I am just getting on the #CausalML bandwagon. Are you already doing it? Any favorite models that you have actually used in production? Please share your thoughts and boost this post! #AI #ML #causal

I am just getting on the #CausalML bandwagon. Are you already doing it? Any favorite models that you have actually used in production? Please share your thoughts and boost this post! #AI #ML #causal

Interested in learning #causalML for policy evaluation/learning?
I consolidated my teaching material from different courses for Master and PhD Economics students. Result:
- 10 slide decks
- 22 R notebooks
See https://github.com/MCKnaus/causalML-teaching and thread below

Causal Christmas Tree Challenge: Draw sth with two potential outcome functions.
I drew a christmas tree and passed the task to my #causalML students @unitübingen@bawü.social as assignment.
Can you eyeball which implied CATE looks like batman? If not, check the notebook https://mcknaus.github.io/assets/cctc/Causal_Christmas_Tree_Challenge2022.html

Hi! #introduction by paper

Do you 👍 #causalML to estimate CATEs? Me too 🤓

With Phillip Heiler, I take a step back and ask what CATEs actually mean if a 0/1 "treatment" is itself heterogeneous (binarized multi trmt or multi trmt versions).

2 new features of a 1yo paper:

You 👍 intution on #causalinference?
😀-based intro to the issue https://t.co/WkDIf9HxyN

You 👍 #econometrics theory?
Updated #DoubleML theory allows no. of trmts -> ∞ -> extreme pscores (limited overlap) https://arxiv.org/abs/2110.01427

Hi! #introduction by paper

Do you 👍 #causalML to estimate CATEs? Me too 🤓

With Phillip Heiler, I take a step back and ask what CATEs actually mean if a 0/1 "treatment" is itself heterogeneous (binarized multi trmt or multi trmt versions).

2 new features of a 1yo paper:

You 👍 intution on #causalinference?
😀-based intro to the issue https://t.co/WkDIf9HxyN

You 👍 #econometrics theory?
Updated #DoubleML theory allows no. of trmts -> ∞ -> extreme pscores (limited overlap) https://arxiv.org/abs/2110.01427