XYZ#2
A causal inference newsletter
News
Lars van der Laan Introduces Adaptive Debiased Machine Learning (ADML)
ADML performs model selection (e.g. Lasso) on top of a doubly-robust estimator. The paper claims improved efficiency for CATE estimators (among others) in the presence of sparsity, at no bias cost. Would be interesting to see how it compares with CATE local projections. Link to the paper and preliminary code.
KDD Workshop on Causal Inference from CausalML and EconML Contributors
While the workshop is past (was on Monday in LA), the list of papers and speakers provides a good overview of the direction of causal inference in the industry: CATE estimation and causal discovery are frequent topics.
Tidier.jl Becomes a Meta Package
Tidier.jl, the R tidyverse port to Julia is now a meta package, re-exporrting TidierData.jl, TidierPlots.jl, TidierCats.jl, TidierDates.jl, and TidierStrings.jl.
Old Reads
Budget-split Testing for A/B Testing in Marketplaces
Cannibalization bias happens when the success of a feature comes at the detriment of another one, for example with budget constraints. Researchers at Linkedin propose splitting subjects and budget into independent buckets and randomizing over buckets. The estimated effects are then scaled up. More details in the paper.
Projecting Causal Estimates (e.g. CATE) on Low Dimensional Spaces
Projecting causal estimates is useful when estimation data is much richer in features than production data. The first stage involves doubly-robust estimation in the spirit of Kennedy (2022), while the second stage is an OLS projection on a low-dimensional space. The main contribution of the paper is to show how to do valid inference on the resulting estimates. The procedure is integrated in R’s grf package.
Bolt Explains Bayesian MMM
The article introduces Marketing Mix Modeling (MMM) and explains how Bolt uses PyMC-Marketing for optimal marketing budget allocation. The article is filled with practical tips and Python code. There could be more details on the trade-offs of a Bayesian approach.
No Consensus from Researchers on CATE Estimators in a Data Challenge
In the 2018 Atlantic (now American) Causal Inference Conferences (ACIC), various researchers were asked to estimate Conditional Average Treatment Effects employing methods of their choice. While there was consensus on the Average Treatment Effect (ATE), CATE estimates differed considerably.
For more causal inference resources:
- Subscribe to my stories on Medium
- Follow me on LinkedIn
- Star
awesome-causal-inference
, where I collect all the articles and additional causal inference resources