Timeline for Conditional independence tests not respecting d-separation
Current License: CC BY-SA 4.0
11 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Dec 21, 2021 at 8:20 | vote | accept | DaSim | ||
| Dec 20, 2021 at 16:29 | comment | added | DaSim | Let us continue this discussion in chat. | |
| Dec 20, 2021 at 16:06 | comment | added | DaSim | I did the same reasoning that you did but I posted a wrong example, sorry. I just modified the question with an updated example that shows the problem in a clearer way | |
| Dec 20, 2021 at 14:45 | history | edited | mribeirodantas | CC BY-SA 4.0 |
added explanations about p-values
|
| Dec 20, 2021 at 14:35 | comment | added | mribeirodantas | I understood what's tricking you. I'm editing my answer now to include it. | |
| Dec 20, 2021 at 13:48 | comment | added | DaSim | My question was, in other words, why I don't see the p-value of the first term greater than the one of the second test? By reading the paper it seems they should since they are using their opposite as dependency function and $dep(B,T|\emptyset) > dep(B,T|A)$ otherwise they should have used something else (like Pearson correlation in the case of linear functions). Many thanks however for the answer, but I think I'll wait a few days since I feel I still have this open point on the implementation with negative p-values that is not totally clear to me | |
| Dec 20, 2021 at 9:19 | comment | added | mribeirodantas | By the way, if you consider your first question answered, please consider accepting my answer as an answer to your question 😁 | |
| Dec 20, 2021 at 9:18 | comment | added | mribeirodantas | You're welcome. I also work with this (though through BN inference). You can check the papers with causal in the title in this link to see how we do that. MIIC, our algorithm, is a constrained-based method for Bayesian network inference, but it's also somewhat score-based in the sense that we make use of scores to pick the contributors that will lead to independence. I'm not sure if I understood your new question. What behaviors? Reflect on p-values in what way? Regarding the paper, I had never heard of "neg p-value" | |
| Dec 20, 2021 at 8:31 | comment | added | DaSim | Thanks a lot for your answer and for the great pointer to Cinelli's answer! I still have a doubt then, that is why these behaviors of mutual information and Pearson correlation do not reflect on the p-values? I have this doubt since I'm implementing this approach and they use as dep function (i.e. a function that grows as conditional independence grows) the negative p-values and this seems to be a common practice (first paragraph of page 4) | |
| Dec 19, 2021 at 13:17 | history | edited | mribeirodantas | CC BY-SA 4.0 |
added 22 characters in body
|
| Dec 19, 2021 at 1:53 | history | answered | mribeirodantas | CC BY-SA 4.0 |