OMG - this short video was so helpful!! Have been struggling with this concept -- reading papers and still confused-- but the figures (especially the plot) and explanations finally made it make sense. I could cry 🥲
Sure thing. The originals are here github.com/NickCH-K/causalgraphs and in general all the code for all these slides is here github.com/NickCH-K/CausalitySlides
I'm describing the pathway X M Y. When describing a pathway it's usually more convenient to list the variables in the order they appear on the path starting from treatment and going to outcome.
All of your videos are great - thank you so much Dr. HK. I've got a question if you've got a minute. Let's say you have 3 variables: 1) minutes of sleep last night 2) caffeine consumed today, and 3) subjective affect today. Both sleep and caffeine cause affect to increase, however observationally people consume more caffeine when they've slept less. Is it correct to say to understand the causal relationship between, separately, 1) caffeine on affect and 2) sleep on affect, I should control for: 1) sleep and 2) caffeine respectively? Thanks so much!
Thanks! It sounds like you're envisioning a diagram with sleep -> caffeine, sleep -> affect, and caffeine -> affect. If that's true, then yes you'd want to control for sleep to get the effect of caffeine, and caffeine to get the effect of sleep not counting how sleep affects caffeine. That said, the effect of sleep controlling for the amount of caffeine would be an example of "mediation analysis" and so "controlling for" isn't quite as simple as just adding a predictor to a regression or something. I'd recommend looking more into mediation analysis for that one.
@@NickHuntingtonKlein Thanks for the response, Nick! I've just ordered your book and intend to use it as accompaniment as I go through your videos more systematically. Is there a page # in there where I can get up-to-speed on mediation analysis?
@@samcialek Thank you! Unfortunately my book doesn't have a mediation section. However you can find a guide in the link at the end of this post. Note the 'Baron and Kenny' approach is the first one it describes, which is the simple approach I said has some issues. ademos.people.uic.edu/Chapter14.html
Thank you, well made. Just an observation about the issue you pointed out on Lecture 17 (Causal Diagrams Practice, slide 6, from you'r website ): B looks like a collider; controlling for D opens a path through D I think. Maybe this is the reason behind daggity's suggestion which states that A is independent from C only if you control for both D and B.
I get the bias which controlling for colliders may cause when we have 3 such variables. But consider your first example with 5 variables where we want to find the effect of X on Y. Suppose X = education Y = earnings A = parents' income and B = parents' education. Let M be the size of the car owned by your parents. Clearly A and B both affect M. Also A affects X, B affects Y. We want to find the effect of X and Y. Here, would we not have to control for A and B? Also, if we controlled for M what additional bias would it create?
In this case we actually don't need to control for A and B (at least on the diagram you're describing. In real life parents education/income would both cause your own too and we'd have arrows from A to Y and B to Y, and we would need to control for A and B). But taking your diagram as given, A and B would be unrelated, so the X M Y pathway doesn't drive an association between X and Y, and doesn't bias the relationship.
@@NickHuntingtonKlein oh yeah, my example is a bit off. But i think I get it. If you could give me a real life example with 5 variables where such a diagram works, it'd be great.
@Ishaan Sengupta the example you gave actually does work in that it tells us not to control for the size of your parents car (it's just in your case it also says you do want to control for A and B as well). If we replace B with something that causes M and X but not Y (say, the price of gas while you're in school maybe? Just as a quick thought) then that tells us not to control for M, and that we don't need to control for B, unless M is controlled for
OMG - this short video was so helpful!! Have been struggling with this concept -- reading papers and still confused-- but the figures (especially the plot) and explanations finally made it make sense. I could cry 🥲
Excellent! Great explanation, laced perfectly with tangible examples
Nick, your videos are of much help! Thank you, greetings from Brazil! :D
Thank you!
Thank you Nick for the video, very well explained!
Thank you Nick! Great explanation and lectures online ;)
Any chance you could share how you created the graphs that visualize controlling for variables (7:32-7:52)?
Sure thing. The originals are here github.com/NickCH-K/causalgraphs and in general all the code for all these slides is here github.com/NickCH-K/CausalitySlides
@@NickHuntingtonKlein Much appreciated!
Thank you. Very videos help so much to understand the concepts.
Can you explain why you read the flow that way at 2:13 ? The arrows appear to be pointing opposite the reading. Thank you!
I'm describing the pathway X M Y. When describing a pathway it's usually more convenient to list the variables in the order they appear on the path starting from treatment and going to outcome.
@@NickHuntingtonKlein thank you!!
Thank you for being clear
omg i did the same analysis a few months ago! awsome!!!
All of your videos are great - thank you so much Dr. HK. I've got a question if you've got a minute. Let's say you have 3 variables: 1) minutes of sleep last night 2) caffeine consumed today, and 3) subjective affect today. Both sleep and caffeine cause affect to increase, however observationally people consume more caffeine when they've slept less. Is it correct to say to understand the causal relationship between, separately, 1) caffeine on affect and 2) sleep on affect, I should control for: 1) sleep and 2) caffeine respectively? Thanks so much!
Thanks! It sounds like you're envisioning a diagram with sleep -> caffeine, sleep -> affect, and caffeine -> affect. If that's true, then yes you'd want to control for sleep to get the effect of caffeine, and caffeine to get the effect of sleep not counting how sleep affects caffeine. That said, the effect of sleep controlling for the amount of caffeine would be an example of "mediation analysis" and so "controlling for" isn't quite as simple as just adding a predictor to a regression or something. I'd recommend looking more into mediation analysis for that one.
@@NickHuntingtonKlein Thanks for the response, Nick! I've just ordered your book and intend to use it as accompaniment as I go through your videos more systematically. Is there a page # in there where I can get up-to-speed on mediation analysis?
@@samcialek Thank you!
Unfortunately my book doesn't have a mediation section. However you can find a guide in the link at the end of this post. Note the 'Baron and Kenny' approach is the first one it describes, which is the simple approach I said has some issues.
ademos.people.uic.edu/Chapter14.html
Thank you, well made.
Just an observation about the issue you pointed out on Lecture 17 (Causal Diagrams Practice, slide 6, from you'r website ):
B looks like a collider; controlling for D opens a path through D I think. Maybe this is the reason behind daggity's suggestion which states that A is independent from C only if you control for both D and B.
Glad you like it! And correct, controlling for B on that slide opens a pathway from A -> B
This is great!
I get the bias which controlling for colliders may cause when we have 3 such variables.
But consider your first example with 5 variables where we want to find the effect of X on Y.
Suppose X = education Y = earnings A = parents' income and B = parents' education. Let M be the size of the car owned by your parents. Clearly A and B both affect M. Also A affects X, B affects Y. We want to find the effect of X and Y.
Here, would we not have to control for A and B?
Also, if we controlled for M what additional bias would it create?
In this case we actually don't need to control for A and B (at least on the diagram you're describing. In real life parents education/income would both cause your own too and we'd have arrows from A to Y and B to Y, and we would need to control for A and B).
But taking your diagram as given, A and B would be unrelated, so the X M Y pathway doesn't drive an association between X and Y, and doesn't bias the relationship.
@@NickHuntingtonKlein oh yeah, my example is a bit off. But i think I get it. If you could give me a real life example with 5 variables where such a diagram works, it'd be great.
@Ishaan Sengupta the example you gave actually does work in that it tells us not to control for the size of your parents car (it's just in your case it also says you do want to control for A and B as well). If we replace B with something that causes M and X but not Y (say, the price of gas while you're in school maybe? Just as a quick thought) then that tells us not to control for M, and that we don't need to control for B, unless M is controlled for
@@NickHuntingtonKlein Thanks a ton, Nick!
god bless you bro