Thank you so much Leslie, I've been looking for this type of video to explain some of the terminology I found in the "Book of Why" by Judea Pearl, it took me forever to find it though. Thank you!!
That's a great book! If you're looking for a deeper dive reading-wise, I recommend "Causal Inference in Statistics: A Primer" (free online at bayes.cs.ucla.edu/PRIMER/) and "Causal Inference: What If" (free online at www.hsph.harvard.edu/miguel-hernan/causal-inference-book/).
Hello, Leslie, great video! I have a question. It seems that the relationship between X and Y should be in opposition if the system is binary. If we know that the value of having a scholarship is yes (Z=1), knowing that the value of musical talent is yes (X=1) indicates that the value of having a veteran parent is no (Y=0). However, we can think of a model of a scholarship in which it accepts not only "only students with musical talent" or "only students with a veteran parent" but also "students with musical talent AND a veteran parent." ([X=1 and Y=1] or [X=1 or Y=1] -->Z=1) In that case, when we fix having a scholarship to yes (Z=1), knowing that musical talent is yes (X=1) does not indicate that having a veteran parent is no (because it may be Y=1). Is this model still a collider? if yes, can we still say that in collider if we know X, we can know Y, given Z conditioned?
Thank you for your question, Banin! In the video, the collider example I give is the one that you describe in your second paragraph: Getting a scholarship (Z=1) occurs if either the student has musical talent (X=1), or the student has a veteran parent (Y=1), or both. You are correct that if we know that Scholarship = Yes, knowing that musical talent = yes does not necessarily indicate that veteran parent is yes (or no)--veteran parent could be either. But the way to think about the induced dependence (conditional dependence) from conditioning on a collision variable is that conditioning on Z can change the conditional probabilities relating X and Y (P(X | Y) and P(Y | X))--just change--not necessarily change to 0 or 1. I think the best way I can explain it is to point you to another video where I discuss this very example in formal probability terms: ua-cam.com/video/acofPsJDq5Y/v-deo.html Hope this helps clarify!
Nice. In the chain example with health insurance, you allude to other causes for utilizing healthcare that are not shown in the diagram and claim that the results of conditioning on utilization covers all effects on the result. I think it's hazardous to teach that claims can be made for partially complete causal diagrams. Diagrams should always be completed before exploring the effects of conditioning.
That's a very fair point! It's a point I make when I'm with my students in person but which I didn't put into the video. I will keep this in mind when making videos in the future.
these videos are all extremely well done. clean recordings, good explanations, thanks for your efforts!!
Great explanation 👌
🎉🎉🎉
Thank you so much Leslie, I've been looking for this type of video to explain some of the terminology I found in the "Book of Why" by Judea Pearl, it took me forever to find it though. Thank you!!
That's a great book! If you're looking for a deeper dive reading-wise, I recommend "Causal Inference in Statistics: A Primer" (free online at bayes.cs.ucla.edu/PRIMER/) and "Causal Inference: What If" (free online at www.hsph.harvard.edu/miguel-hernan/causal-inference-book/).
can you share your slide in the causal inference series?
Clear!!!!!
Thankyou
Hello, Leslie, great video! I have a question. It seems that the relationship between X and Y should be in opposition if the system is binary. If we know that the value of having a scholarship is yes (Z=1), knowing that the value of musical talent is yes (X=1) indicates that the value of having a veteran parent is no (Y=0).
However, we can think of a model of a scholarship in which it accepts not only "only students with musical talent" or "only students with a veteran parent" but also "students with musical talent AND a veteran parent." ([X=1 and Y=1] or [X=1 or Y=1] -->Z=1) In that case, when we fix having a scholarship to yes (Z=1), knowing that musical talent is yes (X=1) does not indicate that having a veteran parent is no (because it may be Y=1).
Is this model still a collider? if yes, can we still say that in collider if we know X, we can know Y, given Z conditioned?
Thank you for your question, Banin! In the video, the collider example I give is the one that you describe in your second paragraph: Getting a scholarship (Z=1) occurs if either the student has musical talent (X=1), or the student has a veteran parent (Y=1), or both. You are correct that if we know that Scholarship = Yes, knowing that musical talent = yes does not necessarily indicate that veteran parent is yes (or no)--veteran parent could be either. But the way to think about the induced dependence (conditional dependence) from conditioning on a collision variable is that conditioning on Z can change the conditional probabilities relating X and Y (P(X | Y) and P(Y | X))--just change--not necessarily change to 0 or 1. I think the best way I can explain it is to point you to another video where I discuss this very example in formal probability terms: ua-cam.com/video/acofPsJDq5Y/v-deo.html
Hope this helps clarify!
@@lesliemyint1865 it seems I missed the "just change" assumption. I will watch the video you refer to. Thank you for answering, really helpful!
Nice. In the chain example with health insurance, you allude to other causes for utilizing healthcare that are not shown in the diagram and claim that the results of conditioning on utilization covers all effects on the result. I think it's hazardous to teach that claims can be made for partially complete causal diagrams. Diagrams should always be completed before exploring the effects of conditioning.
That's a very fair point! It's a point I make when I'm with my students in person but which I didn't put into the video. I will keep this in mind when making videos in the future.