I have a categorical age variable. I’d like to add a random effect for age category, but have those random effects change smoothly with the (ordered) age category, with a random walk or autoregressive prior. Is there an easy way to do this using brms? The closest I could think of is to use mgcv-sty…

After stewing on this for a couple days, my best solution is to sort of “brute force” it using a combination of priors/stanvars to add the code that I need. Here’s an outline of my half-baked solution: Set the model up with random effects for each agecat, i.e. Y ~ (1|agecat) “Remove” the prior on…

Just checking again if anyone has an easier solution? I was hoping there was some sort of built-in functionality in brms that would allow this. @paul.buerkner any ideas?

I don’t think there is a much better way with brms than to what you do currenty. Perhaps except for the fact that I would not set a wide prior on the SD, but rather fix it using something like prior(constant(10000), class = "sd", ...). I don’t fully understand yet why you want to remove the prior on…

Would it make more sense to treat it as a monotonic population effect, and just use the prior to regularize like you normally would see in a random effect? You could also just code this explicitly as a random walk, which is by definition, I think, a function over discrete, but correlated, categorie…

Thanks for the response. I was hoping there would be something easier, but I guess not. Yes, and the constant prior for the SD would be preferred. The reason I want to remove the prior on the random effects is that I’m going to be putting another prior on top of it (manually, as part of step 3 abov…

@bachlaw I don’t want to treat the random effects as monotonic, just smooth. Also I’m looking for a solution using brms… I know how to do this using regular Stan code.

@paul.buerkner - here is an update, and unfortunately I am a bit stuck. I have successfully implemented the random walk prior in Stan (not brms). To do this, I have defined this rw1() function, curtesy of @Bob_Carpenter - real rw1_lpdf(vector alpha, real sigma) { int N = cols(alpha); ret…

I would recommend generating the Stan code via make_stancode, edit as you need (i.e., remove the not needed independent normal prior), fit the model with Stan directly and put all of it back into brms. See the empty = TRUE argument of brm() and corresponding example for details.

Hi! As I ran into the problem of setting up an RW prior in brms. I figured an alternative to the above is possible more straightforward? Here is what I came up with below. This is what I will use as basis for my current project for the moment being (so things may change). Note that you have to scr…

Random walk or autoregressive prior in brms

Modeling

jgellar August 18, 2022, 8:12pm 10

Great solution, thanks.

Topic		Replies	Views
Random walk prior Modeling specification	19	7438	September 4, 2020
Random walk in RStan Modeling rstan , specification , r , prior-choice , priors	11	1337	June 12, 2024
How to add extra prior to brms brms specification	4	1087	April 11, 2022
Using ar() in brms for structured priors - any examples? Modeling brms	5	241	November 8, 2024
How to specify the priors of a random effect ("sds") for a gamm in brms brms specification , brms	9	1317	October 11, 2022

Random walk or autoregressive prior in brms

Related topics