Composing Stan models (posterior as next prior)

General
21

Ah, I see! I knew there was a catch :)
Sure, you can reach me at koenig.chris@gmail.com .

Looking forward to hearing from you!

Best
Chris

22

I have related question. I apologise in adavance if this might be based on misconceptions but is it possible to just use the posterior mean and SD and use them as the mean and SD for a normal prior or other fitting a studentā€™s t distribution and use those for priors? If so, is there a different between regression models with random effects (e.g. subjects), with more than one predictor (e.g. x and x*x or x and z)? Do you need to extract more than just the fixed effects to make this work? This obviously assumes that the respective posterior distributions are well captured by the normal or studentā€™s distribution.

1 Like
23

Hi JAQent,

yes it is. For the scale, however, I would use a measure of precision of the estimate in question. If there are more than one predictors, it boils down to the question of similarity between the previous and current study.

Best
Chris

1 Like
24

Picking up on this again, Iā€™m wondering whether anyone has experience in implementing the Power prior for mixed effects models, where both the current and historical data are modelled using hierarchical models.The method is outlined in section 4 of https://projecteuclid.org/download/pdf_1/euclid.ss/100921267 - it involves specifying a power prior which is the integral of the historical data likelihood (multiplied by the density of the random effects) with respect to the random effects parameters, which I canā€™t figure out how to implement in Stan.

Any thoughts would be much appreciated
Ben

1 Like
25

The link seems broken. Iā€™m up for doing it, if you can open another thread just to keep things on topic.

1 Like
26

Thank you - Power Prior for a Hierarchical Model in Stan

27

Hi everyone,
@bgoodri where do I can find more references for simplex and constrained multivariate distribution?
I am interested in talking in one polytope class that I have.

Thanks

28

I would start with

and

1 Like
29

Thanks was a useful start