Ezequiel Smucler
@esmucler.bsky.social
statistics, machine learning
The following might help: jmlr.org/papers/v21/1..., arxiv.org/abs/2201.02037, academic.oup.com/biomet/artic...
All the different algorithms are implemented in DoWhy too
All the different algorithms are implemented in DoWhy too
Efficient Adjustment Sets for Population Average Causal Treatment Effect Estimation in Graphical Models
jmlr.org
July 9, 2025 at 12:15 PM
The following might help: jmlr.org/papers/v21/1..., arxiv.org/abs/2201.02037, academic.oup.com/biomet/artic...
All the different algorithms are implemented in DoWhy too
All the different algorithms are implemented in DoWhy too
Good question, I’ll DM you
June 21, 2025 at 5:56 PM
Good question, I’ll DM you
The whole thing in the denominator, poor type setting. The usual way to write the score test statistic is indeed with everything squared; it yields the same confidence set in the end
June 21, 2025 at 5:46 PM
The whole thing in the denominator, poor type setting. The usual way to write the score test statistic is indeed with everything squared; it yields the same confidence set in the end
We also contributed an implementation of the method to the excellent DoubleML Python library: docs.doubleml.org/stable/examp...
We believe the score confidence set is an excellent addition to the practitioners toolbox, offering robustness to weak instruments at (asymptotically) no cost.
We believe the score confidence set is an excellent addition to the practitioners toolbox, offering robustness to weak instruments at (asymptotically) no cost.
Python: Confidence Intervals for Instrumental Variables Models That Are Robust to Weak Instruments — DoubleML documentation
docs.doubleml.org
June 21, 2025 at 4:46 PM
We also contributed an implementation of the method to the excellent DoubleML Python library: docs.doubleml.org/stable/examp...
We believe the score confidence set is an excellent addition to the practitioners toolbox, offering robustness to weak instruments at (asymptotically) no cost.
We believe the score confidence set is an excellent addition to the practitioners toolbox, offering robustness to weak instruments at (asymptotically) no cost.
We prove that this confidence set is, in a sense, optimal, achieving the smallest possible diameter and that the standard double machine learning estimator for instrumental variable models is biased under weak instruments.
June 21, 2025 at 4:46 PM
We prove that this confidence set is, in a sense, optimal, achieving the smallest possible diameter and that the standard double machine learning estimator for instrumental variable models is biased under weak instruments.
We study the score confidence set, obtained by inverting the score test constructed from an estimate of the nonparametric influence function for the functional that identifies the LATE. It is known that the score confidence set is robust to weak instruments.
June 21, 2025 at 4:46 PM
We study the score confidence set, obtained by inverting the score test constructed from an estimate of the nonparametric influence function for the functional that identifies the LATE. It is known that the score confidence set is robust to weak instruments.
We study the score confidence set, obtained by inverting the score test constructed from an estimate of the nonparametric influence function for the functional that identifies the LATE. It is known that the score confidence set is robust to weak instruments.
June 21, 2025 at 4:43 PM
We study the score confidence set, obtained by inverting the score test constructed from an estimate of the nonparametric influence function for the functional that identifies the LATE. It is known that the score confidence set is robust to weak instruments.
I agree, as it is written it is just wrong
April 25, 2025 at 4:18 PM
I agree, as it is written it is just wrong
Do you have any references at hand for this?
April 22, 2025 at 8:26 PM
Do you have any references at hand for this?
We also propose a method for constructing uniformly valid confidence sets when all variables are discrete and discuss its (multiple) limitations.
February 25, 2025 at 8:27 AM
We also propose a method for constructing uniformly valid confidence sets when all variables are discrete and discuss its (multiple) limitations.
We establish several impossibility results. In particular, we show that if the model allows the integral equation to be arbitrarily ill-posed, Wald confidence intervals are not asymptotically uniformly valid, and uniformly consistent estimators do not exist.
February 25, 2025 at 8:27 AM
We establish several impossibility results. In particular, we show that if the model allows the integral equation to be arbitrarily ill-posed, Wald confidence intervals are not asymptotically uniformly valid, and uniformly consistent estimators do not exist.
A key feature of these estimands is that they can be expressed as linear functionals of the solution to an integral equation involving conditional expectations.
February 25, 2025 at 8:27 AM
A key feature of these estimands is that they can be expressed as linear functionals of the solution to an integral equation involving conditional expectations.
I for one am not a fan, and resented the treatment of Kolmogorov throughout the book
February 16, 2025 at 7:21 PM
I for one am not a fan, and resented the treatment of Kolmogorov throughout the book
Agree, you should adjust for prognostic variables in all cases
November 26, 2024 at 9:25 PM
Agree, you should adjust for prognostic variables in all cases
The reductions in marginal variance and in conditional (on the observed imbalance) bias that you get from adjusting are two sides of the same coin, this is explained clearly in the chapter of What If reference earlier in this thread
November 26, 2024 at 9:06 PM
The reductions in marginal variance and in conditional (on the observed imbalance) bias that you get from adjusting are two sides of the same coin, this is explained clearly in the chapter of What If reference earlier in this thread
But that's inefficient right? You'll get a lower variance using the pre-treatment information as covariates for regression adjustment (with treatment interactions)
October 30, 2024 at 2:52 PM
But that's inefficient right? You'll get a lower variance using the pre-treatment information as covariates for regression adjustment (with treatment interactions)
Would you use it with experimental data in some scenario?
October 30, 2024 at 2:47 PM
Would you use it with experimental data in some scenario?