Is it possible to update already fitted models adding new covariates? #224
Replies: 2 comments 1 reply
-
Yes, you can use the previous mode as starting point for the "common" parameters. The practical issue is how to either expend the joint theta-vector with the new parameters (if you're lucky they will go at the end of the vector, but I'm not sure that's always the case...), or to extract them and use component-specific initial value settings ( |
Beta Was this translation helpful? Give feedback.
-
Thanks Finn! I see. So using control.mode would require that one makes sure the ordering is correct. Otherwise one can set the inital values in the formula. Best, |
Beta Was this translation helpful? Give feedback.
-
Hi,
I am doing a forward model selection to select covariates (rw2 processes) along with a SPDE. Models are relatively big so I was wondering whether I could keep information from a model with N covariates to ease the model fit of a model with N+1 covariates (N referring t the same covariates).
cmp1 = "y~ Intercept(1) + X1(X1,model="rw2",scale.model=T) + ST(coordinates, model=spde, group...)
cmp2 = "y~ Intercept(1) + X1(X1,model="rw2",scale.model=T) + X2(X2,model="rw2",scale.model=T) + ST(coordinates, model=spde, group...)
I found this https://groups.google.com/g/r-inla-discussion-group/c/Eo2zIEoqTeM/m/Fo1YlHkmAQAJ, which does not do the exact same thing, but it prbably could if we inherit the modes from model 1 "bru1$mode$theta"
bru1$mode$theta
log(Range) for ST log(Stdev) for ST Log precision for X1
3.911919 0.455041 4.592263
and add a "good guess" for the new covariate (X2).
Log precision for X2
2.5
Does this make sense at all?
Cheers
Beta Was this translation helpful? Give feedback.
All reactions