Logo

MINDS & CIS Fall Seminar Series - Shared screen with speaker view
Ilya Shpitser
38:58
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4276366/
Ilya Shpitser
39:01
here’s the paper
Ilya Shpitser
55:05
Comment: a similar confusion occurred historically with ‘structural equation models (SEMs)’ and whether those models are causal. Pearl wrote at length about the history of SEMs. They started off causal with Sewall Wright’s genetics work, but eventually lost that interpretation in many applications, and became just parametric statistical models. Confusion persists to this day.
Ilya Shpitser
01:08:14
It’s worth noting that assuming unobserved confounding is a deterministic function of the treatment logically implies that correlation is causation. (Which means (a) the assumption is too strong, and (b), if you make that assumption just regress the treatment on outcome, and treat the result causally, no need for fancy methods at all.)
Ilya Shpitser
01:21:59
‘unobserved confounding’ is the central problem in causal. its our version of ‘how do we generalize well’ in ML. no one good answer.
Ilya Shpitser
01:27:42
I think this is the survey betsy meant: https://arxiv.org/abs/2009.10982
Soledad Villar
01:27:48
thanks
Soledad Villar
01:28:35
thanks Betsy for a great talk