[FRIAM] Judea Pearl: Book of Why

[uǝlƃ ☣]

Well, the argument I often end up making is that you can do a kind of face validation with the fake data. Show it to someone who's used to dealing with that sort of data and if the fake data looks a lot like the data they normally deal with, then maybe more data-taking isn't necessary. If it looks fake to the "expert", then more data-taking is definitely needed.

On 4/19/20 8:29 AM, Marcus Daniels wrote:
> I have a hard time with this as a way to extend data.   If it is high-dimensional it will be under-sampled.  Seems better to me to  measure or simulate more so that the joint distribution can be realistic.  And if you can do that there is no reason to infer the joint distribution because you *have* it. 
> 
>> On Apr 19, 2020, at 8:18 AM, Frank Wimberly <wimberly3 at gmail.com> wrote:
>>
>> 
>> Going back and forth:  If you infer the causal graph from observational data you can use that graph to simulate data with the same joint distribution as the original data.

-- 
☣ uǝlƃ