[n.b. I’ve posted a similar question with the brms tag inquiring about a brms-specific solution, but I’m still curious if my more-raw-stan-centered proposal below makes any sense] I have space-time data (x,y,time) from the neuroimaging field where the data is collected hierarchically such that ther…

yeah, the example you posted from that repo has a Kronecker product covariance matrix.

Hey! Did you have a look at Rob Tranguccis repo with multi-output GPs ? The code is probably a bit outdated and could perhaps be more efficient. But maybe it’s still helpful to you? Cheers! Max

Good call. I see that the gp_3dim_ncp.stan model specifically seems to implement the isotropy within space and anisotropy between space and time that I’d be looking for (if I’m even using those terms correctly). Very different from how I was thinking of the ARD analogy though, so I’ll have to work o…

Space time models should use a spatiotemporal covariances function. Or a kronecker product structure. Resist the urge to think of it as a 3D space model.

[image] anon75146577: Resist the urge to think of it as a 3D space model. Ah, which I think my ARD-like approach would imply. Gotcha. [image] anon75146577: Or a kronecker product structure From the latest commit message referring to Kronecker products and the code implementing the lat…

@rtrangucci Does your like imply agreement that the code at gp_3dim_ncp.stan implements the Kronecker product approach that Dan suggests? Just wanted to confirm before I dive into adding hierarchy.

I am really interested in spatio-temporal GPS and specifically I am using INLA to infer model parametres about rainfall events. I would really like to use STAN (both HMC or maybe future Laplace approximation when it will be ready, I ping @charlesm93 if there are news) to extend my INLA model(s), tha…

Rob’s code implements the non-hierarchical scenario as Dan suggests it should be done. Presumably you’re looking either for a write up or performance improvements (i.e. you’re happy with the structure of the model and just want it to sample faster)? I think I remember seeing Rob already published on…

Yes, I have a non-hierarchical model scenario and I would like to extend my model with censoring (which is something I cannot do with INLA AFAIK) and have more control over generate quantities in space and time. I never tried the “spline stuff”: have you a link with a simple model to start playing …

[image] fabio: Yes, I have a non-hierarchical model scenario and I would like to extend my model with censoring (which is something I cannot do with INLA AFAIK) INLA has a bunch of survival models, if that helps

Hierarchical 2D+1D space-time modelling using GPs

Modeling

mike-lawrence June 5, 2020, 2:30pm 11

See here for discussion of splines for non-hierarchical space+time models.

Topic		Replies	Views
Running a hierachical spatio-temporal geostaistical model in STAN Modeling specification , spatial	6	1994	April 16, 2020
Space-time models in Stan / BRMS brms ecology	14	4672	August 28, 2019
Expanding the ICAR model with smoothing component to a spatiotemporal framework Modeling rstan	15	1104	January 16, 2025
Using 2D Gaussian process predictions within model Modeling gaussian-process	11	2614	February 18, 2022
Hilbert space GPs with distance matrices Modeling	8	768	July 28, 2023

Hierarchical 2D+1D space-time modelling using GPs

Related topics