search

bambinos / bambi

BAyesian Model-Building Interface (Bambi) in Python.
https://bambinos.github.io/bambi/
MIT License
1.04k stars 119 forks source link

Multivariate Model #108

Open inti opened 6 years ago

inti commented 6 years ago

Hi, Is it possible with bambi to fit multivariate models. I am looking into fitting a model with (at leas) one covariance matrix and two outcomes. I want to estimate the variance components associated with the known covariance (kernel) matrix on each outcome and also the covariance between the outcomes associated with the kernel.

Is that possible to do with bambi

Many thanks in advance

jake-westfall commented 6 years ago

It's not currently possible to fit models like that. If I understood correctly, you seem to be talking about two separate features: (a) fitting models with multiple response variables where you estimate the covariance between responses (/their errors), and (b) fitting models where the realizations of each response are not independent but rather have a certain covariance matrix with parameters to be estimated, as in e.g. serial autocorrelation. (That's my best guess at what you mean by "the variance components associated with the known covariance (kernel) matrix on each outcome.")

Both of these features would be nice to eventually have, at least for Gaussian responses. For non-Gaussian responses I think it would potentially be quite tricky, as you're basically in GEE territory at that point.

inti commented 6 years ago

thanks for the response. your interpretation was correct, sorry if the explanation was a bit confusing. regards, Inti

On 17 November 2017 at 02:05, Jake Westfall notifications@github.com wrote:

It's not currently possible to fit models like that. If I understood correctly, you seem to be talking about two separate features: (a) fitting models with multiple response variables where you estimate the covariance between responses (/their errors), and (b) fitting models where the realizations of each response are not independent but rather have a certain covariance matrix with parameters to be estimated, as in for example GLS https://en.wikipedia.org/wiki/Generalized_least_squares. (That's my best guess at what you mean by "the variance components associated with the known covariance (kernel) matrix on each outcome.")

Both of these features would be nice to eventually have, at least for Gaussian responses. For non-Gaussian responses I think it would potentially be quite tricky, as you're basically in GEE https://en.wikipedia.org/wiki/Generalized_estimating_equation territory at that point.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/bambinos/bambi/issues/108#issuecomment-345146533, or mute the thread https://github.com/notifications/unsubscribe-auth/AAXApdyUW19vzHbsWD6e4p5Ihtrc1GQ4ks5s3RQmgaJpZM4QhEHP .

jake-westfall commented 6 years ago

The title of this issue corresponds to case (a) from above, so I just created #110 to handle case (b).