Functional data analysis (e.g. scalar-on-function regression, as well as functional response types) can be implemented in Stan quite easily - and indeed automated - by using the brms package, which itself calls mgcv to prepare the various matrices. mgcv it turns out can be prodded to do various type…

> brms::make_stancode(octane ~ s(nm, by=NIR, k=50), data=gas) // generated with brms 2.1.0 functions { } data { int<lower=1> N; // total number of observations vector[N] Y; // response variable int<lower=1> K; // number of population-level effects matrix[N, K] X; // population-lev…

Curious! That’s a question for @paul.buerkner .

In brms, the temporary intercept is parameterized to represent the mean of the response distribution, which is roughly 88 in this case. That is we can savely use a weakly informative prior on it when we center around that mean. Such a prior often helps with convergence and we don’t risk “biasing” th…

[image] paul.buerkner: temporary intercept is parameterized to represent the mean of the response distribution I want to understand it a bit more. vector[N] mu = Xc * b + Zs_1_1 * s_1_1 + temp_Intercept; Xc is centered, so I’d expect the mean of the intercept is 0, and Zs_1_1 is the Spl…

I think this goes in a very different direction than what this thread it itended to and has nothing to do with the convergence problems of the model. Please ask these questions again on brms-users ( https://groups.google.com/forum/#!forum/brms-users ) or in a separate thread here on the Stan forums an…

[image] paul.buerkner: I think this goes in a very different direction than what this thread it itended to and has nothing to do with the convergence problems of the model. I’ll respect that answer. However, I don’t agree with the comment. If the model is miss-specified, then of course the …

Sorry, I understand you are trying to help. Let me explain the model. Since the design matrix Xc is centered, the mean of the response distribution is modeled thoroughly via temp_Intercept. The spline terms may change that a bit, but not in a relevant manner here. That is, when removing the prior o…

[image] paul.buerkner: we don’t risk “biasing” the results We have to be careful about what “bias” means here. Bias is just expected estimation error. Or more generally, some effect on the posterior. What happens when you use a data-dependent prior is that there’s usually a small bias from…

Of course you are right Bob. Sometimes I use frequentist wording to shorten what I want to say. That’s some sort of laziness. The idea is that setting such a prior will improve convergence in more complex models, without it’s effect on the posterior being ‘visible’ or at least negigble. The problem…

Scalar-on-function regression using Stan

Modeling

mike-lawrence February 26, 2018, 5:59pm 2

My experience with splines so far is that you indeed need to increase adapt_delta and max_treedepth from their defaults to get good sampling.

Topic		Replies	Views
Using the Apple M1 GPUs: question from a noob brms gpu , splines	46	4520	November 18, 2023
Knots and basis dimension in brms brms	3	2787	November 21, 2019
Restricted cubic splines in Stan? Modeling	19	4215	October 15, 2021
User defined non linear model in brms brms	5	1223	March 25, 2019
Mathemathical notation for a brms model with a smooth term brms notation	8	1265	July 25, 2023

Scalar-on-function regression using Stan

Related topics