Superb introduction for a non-mathematician “domain expert” To understand what the technical expert needs. Unfortunately the underlying quality of the real world data we work with often is insufficiently standardized or machine actionable. This technology is needed for the problems that actually occupy most of a physicians time which is predicting and assessing the effects of treatments particularly once we get off the original guidance from guidelines which might not work in an individual patient.
At the 12:30 mark, X₂←X₁→X₃ is described as a v-structure that can be distinguished from a chain structure with data. That's not a v-structure in that sense, you would need X₂→X₁←X₃.
I have two question and will be grateful for expert and practicioner answers: 1) When calculating CATE you subtract two regressions. This must increase the error considerably. Do we do anything about it? 2) I think in practice, when defining the parameters/independent variables, there's a risk of Simpson paradox. E.g. where's the line between exercising (1) and not exercising (0)? What can one do about it to sleep calmly? Could we do some sort of "hyperparameter tuning" to find the best parameter definitions? It can be tricky...
Question: how do we infer the graphical causal model from data? In the lecture, and the one that follows, we assume a model already exists and use data to answer questions about this model. There are no model selection or model checking involved. Is there a way to infer the causal model from observational data?
It did, I pointed you to the econml documents which describe the many tools to use ML to infer causal models from observational data. There is not one method to explain on UA-cam, there are many, check the work of Susan Athey et al on UA-cam for forest based approaches for one, you can implement those approaches with econml....
DoWhy is about taking a causal graph model, maybe you drew it yourself in daggity, and apply potential outcomes framework to it, which is what you see taught in these lectures.
@@dl5017 I see what you mean now. Afaik Athey still assumes that we know what we are looking for, e.g., how a drug affects clinical outcome. What I was asking is the inference on models themselves, e.g. deciding whether rooster crowing causes sunrise or the other way round from data. The former is on identifucation and estimation, the latter is something else.
OK, replying to myself but also to share with people new to causal inference -- the process of inferring causal graphs from (perhaps observational) data is called "causal discovery".
Brilliant! Thank you for the video and I feel very blessed to be borned in this age when such brilliant lectures are available for free for everyone!
This is the most intuitive and comprehensive guide on causal inference. Thank you Prof. Sontag.
The best lecture on causal inference online
Simply a great lecture. I just recently started diving into this field, and with this lecture I think I have learned the most so far.
Superb introduction for a non-mathematician “domain expert”
To understand what the technical expert needs. Unfortunately the underlying quality of the real world data we work with often is insufficiently standardized or machine actionable. This technology is needed for the problems that actually occupy most of a physicians time which is predicting and assessing the effects of treatments particularly once we get off the original guidance from guidelines which might not work in an individual patient.
Really nice explanations, you kept it simple in the beginning but explained the gist! Thanks for uploading these lectures
What a fantastic teacher and the lecture itself. Thanks for posting this, although I am pretty late to get to it!!
Fantastic explanation! Imma make a video on this topic too.
wow, awesome intro to causal inference!
Great explanation. I wish I had teachers like him.
You do. Right here on UA-cam.
At the 12:30 mark, X₂←X₁→X₃ is described as a v-structure that can be distinguished from a chain structure with data. That's not a v-structure in that sense, you would need X₂→X₁←X₃.
Thank you so much for sharing it with us. It was amazing :)
I have two question and will be grateful for expert and practicioner answers:
1) When calculating CATE you subtract two regressions. This must increase the error considerably. Do we do anything about it?
2) I think in practice, when defining the parameters/independent variables, there's a risk of Simpson paradox. E.g. where's the line between exercising (1) and not exercising (0)? What can one do about it to sleep calmly? Could we do some sort of "hyperparameter tuning" to find the best parameter definitions? It can be tricky...
I would say that Y1 is the red pill and Y0 blue, not the other way
Fantastic!
It's confusing because he doesn't really mention that x is a vector of covariates
Amazing lecture - Thanks!
Where does the counterfactual data come from?
really a great lecture!
Perfect!
Less than 10 minds of the video, I feel like there would be overtreatment of the patient given predictive results. Do more is better than do less
What if a confounder variable only influences the outcome？it's a violation or not
Question: how do we infer the graphical causal model from data? In the lecture, and the one that follows, we assume a model already exists and use data to answer questions about this model. There are no model selection or model checking involved. Is there a way to infer the causal model from observational data?
@@dl5017 Thanks for the suggestions. Although that does not answer my question.
It did, I pointed you to the econml documents which describe the many tools to use ML to infer causal models from observational data. There is not one method to explain on UA-cam, there are many, check the work of Susan Athey et al on UA-cam for forest based approaches for one, you can implement those approaches with econml....
DoWhy is about taking a causal graph model, maybe you drew it yourself in daggity, and apply potential outcomes framework to it, which is what you see taught in these lectures.
@@dl5017 I see what you mean now. Afaik Athey still assumes that we know what we are looking for, e.g., how a drug affects clinical outcome. What I was asking is the inference on models themselves, e.g. deciding whether rooster crowing causes sunrise or the other way round from data. The former is on identifucation and estimation, the latter is something else.
OK, replying to myself but also to share with people new to causal inference -- the process of inferring causal graphs from (perhaps observational) data is called "causal discovery".
Are the problem sets for the course available!? Can't seem to find them
The problem sets are not available for this course. Some instructors do not want to publish them because they are currently in use in the course.
@@mitocw Oh okay. Thanks for the reply!
Nice!
can I translate in arabic (dubbing ) for our students
Yes
@@davidsontag88 even in Chinese, Arabic, or whatever language, I can understand this amazing explanation.... it seems very clear :)
This professor is so handsome.
Bad teaching. Non-coherent. Difficult to follow.
Wonderful!!