Closed drbenvincent closed 1 year ago
I think this is an interesting issue, but not one I know a tonne about...I've heard good things about "Applied Missing Data" by Chris Enders though. Might be able to look into this a bit more after I finish out the Bayesian VAR model thing.
Ok, i've ordered the Enders book - arriving on Friday. I will look into this topic in a bit more detail over Christmas and report back in January if i think i can add anything of interest.
Ok, I think this is definitely something I want to pursue. Think there is a really nice example of workplace empowerment estimation I want to work through.... Will outline a full proposal after I've finished the reliability and prediction pull request if that's alright?
Started some work on this and was able to get FIML and Bayesian imputation working for the multivariate normal. But i had to use a Potential rather than a likelihood as per the discussion here: https://discourse.pymc.io/t/automatic-imputation-of-multivariate-models/11029/3 for the Bayesian MV imputation.
I'm also going to try the chained equation imputation approach which shouldn't need this approach.
Coo! By the way have you seen this video https://www.youtube.com/watch?v=nJ3XefApED0 ?
About a 1/3 of the way through that video
@NathanielF
That video (https://www.youtube.com/watch?v=nJ3XefApED0) needs timestamps, in case you are interested. More info here: https://github.com/pymc-devs/video-timestamps/issues/11
Thanks @reshamas , will have a look tomorrow
I think this is close to done. Really impressed by those jax samplers!! The speed is so much better!
Woop!! Thanks so much @drbenvincent. This one was a real fun one!!
While we do have a lot of example notebooks, we have a distinct lack of examples covering how to deal with missing data. There is also no
missing data
tag.The only ones I can think of are the notebooks on censored and truncated data, which are a form of missing data.
So this is a kind of meta-issue. I/We would be very grateful if people would like to contribute notebooks demonstrating how to handle missing data. Feel free to create specific notebook proposal issues, referencing this issue.