How to predict only using the interaction terms of a variable

Interfaces brms
1

I have a brms model with several variables, I have also a variable, say x interacting with a bunch of other variables (w, y, z). I would like to predict using only x and its interactions, for example the model is

y ~ x*(w + v + z) + a + b +c

I fit the model for all variables and then I just want the value of the posterior of the predictor of x*(a+b+c) , for all my data points, considering that I am only using fixed effects. Reconstructing the interaction data matrix and then multiplying it by the posterior over the coefficients seems like a daunting task.

How can achieve this more straightforwardly in BRMS?

Best regards,

2

You need to pick what values to set the other variables to. Set them to zero? Set them to their means? Average over them in some sort of post-stratified way?

The straightforward thing, and the only thing that really makes sense in the context of your desire to drop the interactions of (a + b + c) with (w + v + z) is to set w, v, and z to zero in the prediction step. Do this by passing newdata to predict.brmsfit()

3

Thank you, I just realised that my question did not really matched my problem, I edited to better reflect my objectives

Sorry for the confusion

4

Cool. Just pass newdata to posterior_linpred, with a, b, and c all zeroed out.

1 Like
5

Hi sorry, for coming back this late, but before closing the issue I just want to be sure of one thing,

if I understand BRMS syntax well the modelwith x*(w+v+z) I will basically be doing (omitting the coefficients):

x + w + v + z + xw + xv + xz, is there a way to get the model that just gives me:

xw + xv + x*z

without having to explicitly write the interactions, as I have many variables. In other words, can I use x*(w+v+z) syntax and be sure that I am only modeling the pairwise interactions?

Thank you again in advance.

6

brms syntax is based on R formula syntax. Check out the : notation; see R: Model Formulae