stefanradev93
commented
2 years ago Hi Michael,
thank you for your interest in our work. I will be happy to read the pre-print when it is available, as this will be one of the few showcases for simulation-based joint modeling!
We have been playing with the idea of estimating hierarchical models for a while and actually want to enable this in the upcoming future branch of BayesFlow. For that, we have started with the "simpler" problem of hierarchical model comparison, for which we will soon have a pre-print and open-source code.
As for the multilevel estimation you describe, there is still work to be done, but we are positive that we will achieve this in the upcoming months. One challenge here is that multilevel neural estimation requires hacky transformations of the neural architectures in order to properly account for all dependencies implied by the underlying graphical model. Stay tuned!
Best, Stefan
mdnunez
bnicenboim
paul-buerkner
Hi Stefan et al.,
Thank you for your research and building this fantastic package. Amin Ghaderi and I are currently building and testing joint models of EEG and behavioural data for each experimental trial. This necessitates approximating posteriors from joint likelihoods on the single-trial level, so we are using BayesFlow. We will likely have a preprint ready within the next couple of weeks.
We are thinking about the future of using BayesFlow for joint modeling (e.g. neural and behavioural data) research. We often use hierarchical models, which we have often implemented using JAGS and Stan etc (https://github.com/mdnunez/pyhddmjags).
Does BayesFlow currently have the ability to fit hierarchical models? And have you thought about his line of research? I suspect it would necessitate some chain of invertible blocks?
A simple hierarchical model extension of the existing DDM example would be a model in which there are many experimental conditions c (e.g. 10). But there is also a hierarchical drift rate parameter, the mean of the condition drift rates:
Likelihood: choice-RT ~ DDM(drift[c], ...) Prior: drift[c] ~ Normal(mu_drift, std_drift^2) Hyperprior: mu_drift ~ Uniform(-7, 7), etc.
This isn't as simple as including an extra normal distribution in the simulator because we want to estimate marginal posterior distributions for the 10 condition drifts: delta[c], as well as marginal posterior distributions for the hierarchical parameters mu_drift and std_drift. I also believe there is a fundamental difference between the model above and this model: Likelihood: choice-RT ~ Complicated_likelihood(mu_drift, std_drift^2, ...) Prior: mu_drift ~ Uniform(-7, 7), etc.
Kind regards,
Michael