I am trying to implement a simulated likelihood model for a stochastic differential equation in stan as described in Simulation and Inference for Stochastic Differential Equations (section 3.3.2). The manual says normal_rng can only be used in generated quantities. All I was able to find is this d…

The reason you can’t use RNGs as part of the model definition is that Stan works by defining a differentiable log density. Using an RNG would break differentiablity as we can’t differentiate through it. It also violates Bayesian information flow principles in the same way as “cut” in BUGS did. If…

Using an RNG would break differentiablity as we can’t differentiate through it. I don’t see why. If you used a constant seed, the random numbers would effectively be constants — you’d have no problem differentiating through them. If you need to generate continuous “random” variables, just defin…

[image] roman: To clarify, the model I am thinking about looks something like this: Let me try another angle. Stan requires you to define a joint log density proportional to the posterior if you want to do full Bayesian inference. Then the information flows from the data to the parameters. …

Sampling doesn’t optimize the log density, but the optimizer does. Yes, at the moment I’m mainly interested in ML/MAP. What you’re suggesting is often done for things like multiple imputation, where it’s too costly to jointly analyze. […] But we don’t support these things in Stan because they’r…

You can generate random numbers in the transformed data block. There’s just not a good way to pass in more general samples other than through data. Even with optimization, we can’t differentiate through the RNG.

What you plan todo sounds very similar to what we have done in our paper: https://www.e-publications.org/ims/submission/AOAS/user/submissionFile/32344?confirm=157ec8b6 We pass in a bunch of random numbers into Stan and we also give recipes to ensure convergence of results. The supplements contain …

I am very interested in reading your supplements of the source code, but I couldn’t open it from the journal website. Do you mind sharing with me? Thanks!

I got access to download the supplements now; could you point out where you pass in the random numbers? The code is pretty long … Thanks!

Cool that someone is diving into this! I recall that the simulation example is straightforward. So start with that. I can dig up the code tomorrow and have a look in case you are still struggling. Note that I generated the random numbers within R if my memory serves me right. EDIT: Now I remember, …

I checked the code of linear_baad.R and pkpd_baad.R, but I am not sure which lines of code did you refer to. I saw the chain id was fixed at 4 in both files.

Generating random numbers in the model

Modeling

rbecerril July 23, 2021, 7:06pm 21

Thanks @mike-lawrence . I just did that. The new post is here: Implementing two-step estimation

1 Like

Topic		Replies	Views
Sampling from a matrix in STAN Modeling rstan , techniques , performance	28	1530	September 1, 2021
How to realize the R code in Stan Modeling rstan	3	801	July 31, 2021
Statistics on random numbers in model block Modeling	4	844	August 8, 2017
Generating random numbers Modeling rstan	10	1185	February 12, 2022
How to generate Random Number for User defined model Modeling fitting-issues	6	749	June 19, 2022

Generating random numbers in the model

Related topics