Comparing multiple models performance to spatially correlated models

Hi all,
I’m trying to investigate if spatially correlated models are better in fitting the data compared to non correlated models.
If I got the correlated model formula correct I would do something like

corr_model <- bf(mvbind(Y1,Y2,Y3,Y4)~ X1 + X2 + X3) + set_rescore(TRUE)

and then fit the model through the brms interface.
Now, I need to compare these models to the models:

m1 <- bf(Y1~ X1 + X2 + X3)
m2 <- bf(Y2~ X1 + X2 + X3)
m3 <- bf(Y3~ X1 + X2 + X3)
m4 <- bf(Y4~ X1 + X2 + X3)

Is there a way to do for example bayes_factor between the corr_model and all the other models? or is there a better way that i could use to formulate the problem?

Thanks in advance.

Have you considered bridge sampling or LOO-CV?

1 Like

If you have spatially autocorrelated data (Exploratory spatial data analysis • geostan), you’ll also want to look at some other options if you haven’t already (such as CAR or SAR models). Otherwise, @dankehila’s comment is a good starting point for this discussion, in addition to examining the degree of correlation in the multivariate model, which might provide reason to use it.

1 Like