Yes I am, great catch! Thanks Jaime That’s a bummer because it ruins this example since X does not block that path. Let’s just imagine the DAG here has Z2->Z4 instead of Z4->Z2 The blog has been updated version of this Front Door Criterion Example: towardsdatascience.com/causal-effects-via-dags-801df31da794?sk=aa0947ca29e23fb3c1612e40deac38cf
I know in your Causal Discovery video you explained how to find a causal model using data alone. And to find Causal Inferences, you have to generate an estimand. However, while using these techniques in a social research, can we determine a DAG on our own hypothesis? Or using other qualitative observational data? Can the DAG be purely human made?
Regarding 11:30 - I don't quite understand why the path is blocked by X? Wouldn't it only be blocked by X if we would condition on it (or if it was a collider)?
Good question. The path from Z1 to Y is blocked by X because X is part of a chain that satisfies the first version of blocking from the definition below. A path p is said to be blocked by a set of nodes {Z_i} if and only if, 1. p contains a chain A -> B -> C or a fork A C, such that B is an element in {Z_i} - This is what we might intuitively think of as blocking 2. p contains a collider (i.e. an inverted fork) A -> B
Great question. This comes down to the definition of a back-door path. Which is any path starting with an arrow point to X and ending with an arrow pointing to Y. All other arrowheads are irrelevant.
@@ShawhinTalebi Thanks, Shawhin Note for myself: As long as it's a path with arrow into X and Y, the direction of any arrows existing between them doesn't matter.
Hello! I have a reference saying this: "If the causal graph doesn’t contain cycles but the noise terms are dependent, then the model is semi-Markovian. ... Finally, the graphs of non-Markovian models contain cycles." May I clarify in 4:00 if you meant Semi-Markovian? Thanks a lot. I just have so many questions cause I'm really confused with all my readings, so I'm relying on your videos for simplification.
Good questions. There are many terms here and it can be a lot to unpack. The models at 4:00 are "Not Markovian", meaning they do not satisfy the following definition: graph has no cycles and noise terms are independent. If a graph has no cycle, but one or more noise terms NOT independent, then the model is said to be "Semi-Markovian". Hope that helps!
More in this series 👇 Intro to Causal Effects: ruclips.net/video/BOPOX_mTS0g/видео.html Propensity Scores: ruclips.net/video/dm-BWjyYQpw/видео.html Do-operator: ruclips.net/video/dejZzJIZdow/видео.html Regression techniques: ruclips.net/video/O72uByJlnMw/видео.html Intro to Causality: ruclips.net/video/WqASiuM4a-A/видео.html Causal Inference: ruclips.net/video/PFBI-ZfV5rs/видео.html Causal Discovery: ruclips.net/video/tufdEUSjmNI/видео.html
In 11:20 , when listing the back door paths from Z2 to Y. Aren't you missing the path Z2 Y?
Yes I am, great catch! Thanks Jaime
That’s a bummer because it ruins this example since X does not block that path.
Let’s just imagine the DAG here has Z2->Z4 instead of Z4->Z2
The blog has been updated version of this Front Door Criterion Example: towardsdatascience.com/causal-effects-via-dags-801df31da794?sk=aa0947ca29e23fb3c1612e40deac38cf
@@ShawhinTalebi now it works :) great videos Shawhin!
Yet another great video! Thank you! 😊
Thanks :)
This kicks ass. Thanks!
I know in your Causal Discovery video you explained how to find a causal model using data alone. And to find Causal Inferences, you have to generate an estimand. However, while using these techniques in a social research, can we determine a DAG on our own hypothesis? Or using other qualitative observational data? Can the DAG be purely human made?
Good question! IMO causal discovery is still quite experimental. Therefore the best way to construct a DAG is via domain expertise.
Great content ❤
Thanks, glad you like it 😁
Regarding 11:30 - I don't quite understand why the path is blocked by X? Wouldn't it only be blocked by X if we would condition on it (or if it was a collider)?
Good question. The path from Z1 to Y is blocked by X because X is part of a chain that satisfies the first version of blocking from the definition below.
A path p is said to be blocked by a set of nodes {Z_i} if and only if,
1. p contains a chain A -> B -> C or a fork A C, such that B is an element in {Z_i} - This is what we might intuitively think of as blocking
2. p contains a collider (i.e. an inverted fork) A -> B
5:01 Why is "X ← Z1 → Z3 ← Z2→ Y" also a back door path, as Z3 doesn't point to Z2?
Great question. This comes down to the definition of a back-door path. Which is any path starting with an arrow point to X and ending with an arrow pointing to Y. All other arrowheads are irrelevant.
@@ShawhinTalebi Thanks, Shawhin
Note for myself:
As long as it's a path with arrow into X and Y, the direction of any arrows existing between them doesn't matter.
An idea for future video. On applying CI to time series data.
Great suggestion! Thanks
Hello! I have a reference saying this: "If the causal graph doesn’t contain cycles but the noise terms are dependent, then the model is semi-Markovian. ... Finally, the graphs of non-Markovian models contain cycles." May I clarify in 4:00 if you meant Semi-Markovian? Thanks a lot. I just have so many questions cause I'm really confused with all my readings, so I'm relying on your videos for simplification.
Good questions. There are many terms here and it can be a lot to unpack.
The models at 4:00 are "Not Markovian", meaning they do not satisfy the following definition: graph has no cycles and noise terms are independent.
If a graph has no cycle, but one or more noise terms NOT independent, then the model is said to be "Semi-Markovian".
Hope that helps!
@@ShawhinTalebi Thanks for the clarification. It surely helped a lot. ❤
More in this series 👇
Intro to Causal Effects: ruclips.net/video/BOPOX_mTS0g/видео.html
Propensity Scores: ruclips.net/video/dm-BWjyYQpw/видео.html
Do-operator: ruclips.net/video/dejZzJIZdow/видео.html
Regression techniques: ruclips.net/video/O72uByJlnMw/видео.html
Intro to Causality: ruclips.net/video/WqASiuM4a-A/видео.html
Causal Inference: ruclips.net/video/PFBI-ZfV5rs/видео.html
Causal Discovery: ruclips.net/video/tufdEUSjmNI/видео.html