1/ Sharing an interesting observation from Frank Wiltzeck's book "Fundamentals."
In the 17th Century, while the entire scientific
world was pre-occupied with planetary motion and other
grand questions of philosophy, Galileo made careful studies of simple forms of motion, e.g.,
2/ how balls roll down an inclined plane and how pendulum oscillate. To most of Galileo's contemporaries such measurements must have appeared trivial, if not irrelevant, to their speculations on how the world works. Yet Galileo aspired to a different kind of understanding.
3/ He wanted to understand something precisely, rather than everything vaguely. Ironically, it was Galileo's type of understanding that enabled Newton's theory of gravitation to explain "how the world works".
What do I mention it? Because we have had lengthy
discussions here
4/4 about the virtues of "toy problems," and
the contempt with which economists (and others) view such problems (see ucla.in/36EoNzO). I still maintain that the kind of understanding gained by solving our "toy problems" is essential for next generation causal inference.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
and "oppressors” and by the way it associates "whiteness" with "oppression" and "colonialism".
I am a "white" Jewish American, and I believe that the history of my people is a model of emancipation from oppression and colonialism, culminating in the State of Israel which is 2/4
an inspirational model of an oppressed ethnic minority lifting itself from the margin of history to become a world center of art, science and entrepreneurship -- a multi-colored light-house of free speech and gender equality.
This question annoys ALL students (and professors) of ML, but they are afraid to ask. Thanks for raising it in this "no hand waving" forum. Take two causal diagrams:
X-->Y and X<--Y, and ask a neural network to decide which is more probable, after seeing 10 billion samples. 1/n
The answer will be: No difference; each diagram scores the same fit as the other. Let's be more sophisticated: assign each diagram a prior and run a Bayesian analysis on the samples. Lo and Behold, the posteriors will equal to the priors no matter how we start. How come? 2/n
Isn't a neural network supposed to learn the truth given enough data? Ans. No! Learning only occurs when the learnable offends the data less than its competitors. Our two diagrams never offend any data, so nothing is learnable. Aha! But what if our data involves interventions? 3/
When I see a paper on explainability, first question I
ask is: "What does it explain?", the data-fitting strategy of
the fitter? or real-life events such as death or survival.
I believe this paper arxiv.org/pdf/2010.10596…
is mostly about the former, as can be seen from the 1/
equations and from the absence of any world-model.
While it is sometimes useful to explain the data-fitting system (eg. for debugging), it is also important to distinguish this kind of counterfactual explanations
from the kind generated in the causal inference literature.
2/3
Beware, a model-blind system might conclude that
the rooster crow explains the sunrise. It might also explain that your loan was denied because you are a male, and would also have been denied if you were a female. I wonder how ML folks would debug this system.
1/4 Comments on your Front-Door paper:
* The expression "a single, strictly exogenous mediator
variable" is problematic: (1) Causality p. 82 defines
FDC as "A set of variables", not "a single variable". (2)
"exogenous mediator" is an oxymoron. I originally
called it (1973):
2/ "Mediating Instrumental Variables" ucla.in/2pJzGNK, best described as an "exogenously-disturbed mediator".
* "The first application of FDC" sounds too pessimistic. Situations involving exogenously-disturbed mediators are at least as plausible as "exclusion-restricted
3/ exogenous variables" (traditional IV's) which were introduced 70 yrs earlier, when DAGs were not
around to invite scrutiny. Imbens comments
reflect that absence ucla.in/36EoNzO
* Why introduce FDC in the context of linear regression
where ATE is identifiable by
1/ I'm glad, Sean, that our brief exchange has resulted in your great clarification of the issues, from which I have learned a lot. Two thoughts come immediately to mind: (1) It is a blessing that we can enjoy a division of labor between CI and statistics, the former generates
2/ causal estimands, the latter estimate them. Note though that the former is not totally oblivious to the type of data available. Different types of data will result in different estimands. eg.,experimental vs. observational, corrupted by missingness or by proxies or by
3/ differential selection etc. (2) I don't buy the mystification of "collecting adequate data". I am in the business of automating a scientist, so, if there is human judgement involved in the data collection task, we do not stop here and surrender it to humans. We see it as an
1/4 In view of the dominant role that re-weighing plays in
extrapolating effects across populations, and the many
Twitter requests for a concise graphical criterion
that gives re-weighing its legitimacy, I am retweeting
the criterion (called "S-admissibility"), in next 4 tweets.
2/ It works on a selection diagram in which S nodes represent disparities between the target (*) population and study population (experimental). Z is a set of measurements. To test if Z is S-admissible (1) Remove all arrows pointing to X (2) Check if {X, Z} d-separates S from Y
3/4 If Z passes this test, then the reweighing formula is valid:
P*(y|do(x)) = SUM_z P(y|do(x),z)P*(z)
In words: Effect at target equals the Z-specific effects at study, averaged over Z, using the target distribution P*(z) as weight.
Warning, this is merely a sufficient test.