Edward H. Kennedy
@edwardhkennedy.bsky.social
assoc prof of statistics & data science at Carnegie Mellon
https://www.ehkennedy.com/
interested in causality, machine learning, nonparametrics, public policy, etc
https://www.ehkennedy.com/
interested in causality, machine learning, nonparametrics, public policy, etc
Once you have a pathwise differentiable parameter, a natural estimator is a debiased plug-in, which subtracts off the avg of estimated influence fn
Pfanzagl gives this 1-step estimator here - in causal inference this is exactly the doubly robust / DML estimator you know & love!
Pfanzagl gives this 1-step estimator here - in causal inference this is exactly the doubly robust / DML estimator you know & love!
February 17, 2025 at 2:47 AM
Once you have a pathwise differentiable parameter, a natural estimator is a debiased plug-in, which subtracts off the avg of estimated influence fn
Pfanzagl gives this 1-step estimator here - in causal inference this is exactly the doubly robust / DML estimator you know & love!
Pfanzagl gives this 1-step estimator here - in causal inference this is exactly the doubly robust / DML estimator you know & love!
Here’s Pfanzagl on the gradient of a functional/parameter, aka derivative term in a von Mises expansion, aka influence function, aka Neyman-orthogonal score
Richard von Mises first characterized smoothness this way for stats in the 30s/40s! eg:
projecteuclid.org/journals/ann...
Richard von Mises first characterized smoothness this way for stats in the 30s/40s! eg:
projecteuclid.org/journals/ann...
February 17, 2025 at 2:47 AM
Here’s Pfanzagl on the gradient of a functional/parameter, aka derivative term in a von Mises expansion, aka influence function, aka Neyman-orthogonal score
Richard von Mises first characterized smoothness this way for stats in the 30s/40s! eg:
projecteuclid.org/journals/ann...
Richard von Mises first characterized smoothness this way for stats in the 30s/40s! eg:
projecteuclid.org/journals/ann...
"Randomized trials should be used to answer any causal question that can be so studied...
But the reality is that observational methods are used everyday to answer pressing causal questions that cannot be studied in randomized trials."
- Jamie Robins, 2002
tinyurl.com/4yuxfxes
tinyurl.com/zncp39mr
But the reality is that observational methods are used everyday to answer pressing causal questions that cannot be studied in randomized trials."
- Jamie Robins, 2002
tinyurl.com/4yuxfxes
tinyurl.com/zncp39mr
January 13, 2025 at 2:49 AM
"Randomized trials should be used to answer any causal question that can be so studied...
But the reality is that observational methods are used everyday to answer pressing causal questions that cannot be studied in randomized trials."
- Jamie Robins, 2002
tinyurl.com/4yuxfxes
tinyurl.com/zncp39mr
But the reality is that observational methods are used everyday to answer pressing causal questions that cannot be studied in randomized trials."
- Jamie Robins, 2002
tinyurl.com/4yuxfxes
tinyurl.com/zncp39mr
Should we use structure-agnostic (arxiv.org/abs/2305.04116) or smooth (arxiv.org/pdf/1512.02174) models for causal inference?
Why not both?
Here we propose novel hybrid smooth+agnostic model, give minimax rates, & new optimal methods
arxiv.org/pdf/2405.08525
-> fast rates under weaker conditions
Why not both?
Here we propose novel hybrid smooth+agnostic model, give minimax rates, & new optimal methods
arxiv.org/pdf/2405.08525
-> fast rates under weaker conditions
December 13, 2024 at 4:07 AM
Should we use structure-agnostic (arxiv.org/abs/2305.04116) or smooth (arxiv.org/pdf/1512.02174) models for causal inference?
Why not both?
Here we propose novel hybrid smooth+agnostic model, give minimax rates, & new optimal methods
arxiv.org/pdf/2405.08525
-> fast rates under weaker conditions
Why not both?
Here we propose novel hybrid smooth+agnostic model, give minimax rates, & new optimal methods
arxiv.org/pdf/2405.08525
-> fast rates under weaker conditions
New paper! arxiv.org/pdf/2411.14285
Led by amazing postdoc Alex Levis: www.awlevis.com/about/
We show causal effects of new "soft" interventions are less sensitive to unmeasured confounding
& study which effects are *least* sensitive to confounding -> makes new connections to optimal transport
Led by amazing postdoc Alex Levis: www.awlevis.com/about/
We show causal effects of new "soft" interventions are less sensitive to unmeasured confounding
& study which effects are *least* sensitive to confounding -> makes new connections to optimal transport
November 22, 2024 at 4:39 AM
New paper! arxiv.org/pdf/2411.14285
Led by amazing postdoc Alex Levis: www.awlevis.com/about/
We show causal effects of new "soft" interventions are less sensitive to unmeasured confounding
& study which effects are *least* sensitive to confounding -> makes new connections to optimal transport
Led by amazing postdoc Alex Levis: www.awlevis.com/about/
We show causal effects of new "soft" interventions are less sensitive to unmeasured confounding
& study which effects are *least* sensitive to confounding -> makes new connections to optimal transport
In this paper we consider incremental effects of continuous exposures:
arxiv.org/abs/2409.11967
i.e., soft interventions on cts treatments like dose, duration, frequency
it turns out exponential tilts preserve all nice properties of incremental effects with binary trt (arxiv.org/abs/1704.00211)
arxiv.org/abs/2409.11967
i.e., soft interventions on cts treatments like dose, duration, frequency
it turns out exponential tilts preserve all nice properties of incremental effects with binary trt (arxiv.org/abs/1704.00211)
November 13, 2024 at 4:26 AM
In this paper we consider incremental effects of continuous exposures:
arxiv.org/abs/2409.11967
i.e., soft interventions on cts treatments like dose, duration, frequency
it turns out exponential tilts preserve all nice properties of incremental effects with binary trt (arxiv.org/abs/1704.00211)
arxiv.org/abs/2409.11967
i.e., soft interventions on cts treatments like dose, duration, frequency
it turns out exponential tilts preserve all nice properties of incremental effects with binary trt (arxiv.org/abs/1704.00211)
Very excited about this paper!
arxiv.org/abs/2305.04116
We study if one can improve popular semiparametric / doubly robust / DML causal effect estimators - w/o adding structural assumptions...
Short answer: nope!
Turns out these methods are minimax optimal here
www.ehkennedy.com/uploads/5/8/...
arxiv.org/abs/2305.04116
We study if one can improve popular semiparametric / doubly robust / DML causal effect estimators - w/o adding structural assumptions...
Short answer: nope!
Turns out these methods are minimax optimal here
www.ehkennedy.com/uploads/5/8/...
November 12, 2024 at 3:30 AM
Very excited about this paper!
arxiv.org/abs/2305.04116
We study if one can improve popular semiparametric / doubly robust / DML causal effect estimators - w/o adding structural assumptions...
Short answer: nope!
Turns out these methods are minimax optimal here
www.ehkennedy.com/uploads/5/8/...
arxiv.org/abs/2305.04116
We study if one can improve popular semiparametric / doubly robust / DML causal effect estimators - w/o adding structural assumptions...
Short answer: nope!
Turns out these methods are minimax optimal here
www.ehkennedy.com/uploads/5/8/...
there are *surprisingly many* open problems when it comes to theory/methods in causal inference
check out this talk by Siva Balakrishnan for an excellent & comprehensive summary of the state of the art
www.youtube.com/live/Mnum0Ox...
www.stat.cmu.edu/~siva/
check out this talk by Siva Balakrishnan for an excellent & comprehensive summary of the state of the art
www.youtube.com/live/Mnum0Ox...
www.stat.cmu.edu/~siva/
November 12, 2024 at 3:15 AM
there are *surprisingly many* open problems when it comes to theory/methods in causal inference
check out this talk by Siva Balakrishnan for an excellent & comprehensive summary of the state of the art
www.youtube.com/live/Mnum0Ox...
www.stat.cmu.edu/~siva/
check out this talk by Siva Balakrishnan for an excellent & comprehensive summary of the state of the art
www.youtube.com/live/Mnum0Ox...
www.stat.cmu.edu/~siva/
November 12, 2024 at 3:10 AM
Short story - the ideas behind “causal ML” and “double machine learning” go back at least 40 years
Here is an estimator from a 1982 textbook that today would be called double machine learning or something similar
Here is an estimator from a 1982 textbook that today would be called double machine learning or something similar
September 30, 2024 at 3:32 AM
Short story - the ideas behind “causal ML” and “double machine learning” go back at least 40 years
Here is an estimator from a 1982 textbook that today would be called double machine learning or something similar
Here is an estimator from a 1982 textbook that today would be called double machine learning or something similar
Pfanzagl also points out our prized n^(1/4) nuisance rate conditions, which allow fast rates & inference for the parameter of interest, even w/ flexible np smoothing / lasso etc
Even throws in a mention of sample-splitting / cross-fitting! (from Hasminskii & Ibragimov 1979)
Even throws in a mention of sample-splitting / cross-fitting! (from Hasminskii & Ibragimov 1979)
September 30, 2024 at 3:27 AM
Pfanzagl also points out our prized n^(1/4) nuisance rate conditions, which allow fast rates & inference for the parameter of interest, even w/ flexible np smoothing / lasso etc
Even throws in a mention of sample-splitting / cross-fitting! (from Hasminskii & Ibragimov 1979)
Even throws in a mention of sample-splitting / cross-fitting! (from Hasminskii & Ibragimov 1979)
Once you have a pathwise differentiable parameter, a natural estimator is a debiased plug-in, which subtracts off the avg of estimated influence fn
Pfanzagl gives this 1-step estimator here - in causal inference this is exactly the doubly robust / DML estimator you know & love!
Pfanzagl gives this 1-step estimator here - in causal inference this is exactly the doubly robust / DML estimator you know & love!
September 30, 2024 at 3:26 AM
Once you have a pathwise differentiable parameter, a natural estimator is a debiased plug-in, which subtracts off the avg of estimated influence fn
Pfanzagl gives this 1-step estimator here - in causal inference this is exactly the doubly robust / DML estimator you know & love!
Pfanzagl gives this 1-step estimator here - in causal inference this is exactly the doubly robust / DML estimator you know & love!
Pfanzagl uses pathwise differentiability above, but w/regularity conditions this is just a distributional Taylor expansion, which is easier to think about
I note this in my tutorial here:
x.com/edwardhkenne...
Also v related to neyman orthogonality - worth a separate thread
I note this in my tutorial here:
x.com/edwardhkenne...
Also v related to neyman orthogonality - worth a separate thread
September 30, 2024 at 3:22 AM
Pfanzagl uses pathwise differentiability above, but w/regularity conditions this is just a distributional Taylor expansion, which is easier to think about
I note this in my tutorial here:
x.com/edwardhkenne...
Also v related to neyman orthogonality - worth a separate thread
I note this in my tutorial here:
x.com/edwardhkenne...
Also v related to neyman orthogonality - worth a separate thread
Here’s Pfanzagl on the gradient of a functional/parameter, aka derivative term in a von Mises expansion, aka influence function, aka Neyman-orthogonal score
Richard von Mises first characterized smoothness this way for stats in the 30s/40s! eg:
projecteuclid.org/euclid.aoms/...
Richard von Mises first characterized smoothness this way for stats in the 30s/40s! eg:
projecteuclid.org/euclid.aoms/...
September 30, 2024 at 3:18 AM
Here’s Pfanzagl on the gradient of a functional/parameter, aka derivative term in a von Mises expansion, aka influence function, aka Neyman-orthogonal score
Richard von Mises first characterized smoothness this way for stats in the 30s/40s! eg:
projecteuclid.org/euclid.aoms/...
Richard von Mises first characterized smoothness this way for stats in the 30s/40s! eg:
projecteuclid.org/euclid.aoms/...
Here’s a (now 35 yr old) book review where Bickel gives some history:
projecteuclid.org/euclid.aos/1...
A more detailed discussion of the early days is in the comprehensive semiparametrics text by BKRW:
springer.com/gp/book/9780...
Would love to hear if others know good historical resources!
projecteuclid.org/euclid.aos/1...
A more detailed discussion of the early days is in the comprehensive semiparametrics text by BKRW:
springer.com/gp/book/9780...
Would love to hear if others know good historical resources!
September 30, 2024 at 3:13 AM
Here’s a (now 35 yr old) book review where Bickel gives some history:
projecteuclid.org/euclid.aos/1...
A more detailed discussion of the early days is in the comprehensive semiparametrics text by BKRW:
springer.com/gp/book/9780...
Would love to hear if others know good historical resources!
projecteuclid.org/euclid.aos/1...
A more detailed discussion of the early days is in the comprehensive semiparametrics text by BKRW:
springer.com/gp/book/9780...
Would love to hear if others know good historical resources!
Some amazing quotes by Herbert Robbins here:
jiayinggu.weebly.com/uploads/3/8/...
"Why does it take so long? Why haven't I done ten times as much as I have? Why do I bother over & over again trying the wrong way when the right way was staring me in the face all the time? I don't know."
jiayinggu.weebly.com/uploads/3/8/...
"Why does it take so long? Why haven't I done ten times as much as I have? Why do I bother over & over again trying the wrong way when the right way was staring me in the face all the time? I don't know."
December 13, 2023 at 3:58 AM
Some amazing quotes by Herbert Robbins here:
jiayinggu.weebly.com/uploads/3/8/...
"Why does it take so long? Why haven't I done ten times as much as I have? Why do I bother over & over again trying the wrong way when the right way was staring me in the face all the time? I don't know."
jiayinggu.weebly.com/uploads/3/8/...
"Why does it take so long? Why haven't I done ten times as much as I have? Why do I bother over & over again trying the wrong way when the right way was staring me in the face all the time? I don't know."
November 5, 2023 at 2:37 AM
Check out this great paper by Mateo Rubio (scholar.google.com/citations?us...), rigorously estimating causal effects of the "cycle of violence"
Superb example of how to tell a story including average effects, heterogeneous effects, & sensitivity analysis (i.e., relaxing assumptions abt confounding)
Superb example of how to tell a story including average effects, heterogeneous effects, & sensitivity analysis (i.e., relaxing assumptions abt confounding)
October 21, 2023 at 2:55 AM
Check out this great paper by Mateo Rubio (scholar.google.com/citations?us...), rigorously estimating causal effects of the "cycle of violence"
Superb example of how to tell a story including average effects, heterogeneous effects, & sensitivity analysis (i.e., relaxing assumptions abt confounding)
Superb example of how to tell a story including average effects, heterogeneous effects, & sensitivity analysis (i.e., relaxing assumptions abt confounding)
October 21, 2023 at 12:00 AM
October 17, 2023 at 4:05 PM
"There are two types of statisticians: those who do causal inference and those who lie about it."
- Larry Wasserman #statsquotes
www.jstor.org/stable/26699...
- Larry Wasserman #statsquotes
www.jstor.org/stable/26699...
October 16, 2023 at 1:42 AM
"There are two types of statisticians: those who do causal inference and those who lie about it."
- Larry Wasserman #statsquotes
www.jstor.org/stable/26699...
- Larry Wasserman #statsquotes
www.jstor.org/stable/26699...
Matteo Bonvini made an R package for our “proportion of unmeasured confounding” sensitivity approach here:
github.com/matteobonvin...
Paper:
arxiv.org/abs/1912.02793
github.com/matteobonvin...
Paper:
arxiv.org/abs/1912.02793
October 14, 2023 at 11:21 PM
Matteo Bonvini made an R package for our “proportion of unmeasured confounding” sensitivity approach here:
github.com/matteobonvin...
Paper:
arxiv.org/abs/1912.02793
github.com/matteobonvin...
Paper:
arxiv.org/abs/1912.02793
Yes X is confounders and Y1 = Y^{a=1} is the potential outcome.
There are lots of different kinds of sensitivity analysis models - the Cinelli & Hazlett one is more specialized/structured, & assumes more.
The lit review in this paper might be useful:
arxiv.org/pdf/2104.083...
There are lots of different kinds of sensitivity analysis models - the Cinelli & Hazlett one is more specialized/structured, & assumes more.
The lit review in this paper might be useful:
arxiv.org/pdf/2104.083...
October 14, 2023 at 11:20 AM
Yes X is confounders and Y1 = Y^{a=1} is the potential outcome.
There are lots of different kinds of sensitivity analysis models - the Cinelli & Hazlett one is more specialized/structured, & assumes more.
The lit review in this paper might be useful:
arxiv.org/pdf/2104.083...
There are lots of different kinds of sensitivity analysis models - the Cinelli & Hazlett one is more specialized/structured, & assumes more.
The lit review in this paper might be useful:
arxiv.org/pdf/2104.083...