I am trying to adapt my model to gamm (which I was using in lmer).
My former formula was
b <- lmer(metric1~a+b+c+d+e+f+g+h+i+(1|X) + (1|Y), data = dataset)
I am trying to adapt this formula to modelling with gamm. But I could not figure out how yet. My purpose is to find the simplified model.
Any ideas or examples would me much appreciated.
If I am not wrong, you are fitting a linear mixed model with two i.i.d. random effects (on intercept).
There is no need to use gamm in this case. Using gam with method = REML will do.
gam(metric1~a+b+c+d+e+f+g+h+i+ s(X, bs = 're') + s(Y, bs = 're'), data = dataset, method = 'REML')
Note I have not extended other fixed effects into smooth functions, which you may do it yourself.
If you have a large dataset, it is suggested to use bam, and note it is method = 'fREML in this case.
The difference between gam and bam in REML estimation, is that the former uses "outer" iteration, while the latter uses "performance" iteration. But for Gaussian data there's no difference, though bam itself is designed for large dataset, using iterative QR reduction and parallel computing on request.
Personally I think gamm is outdated. It does REML estimation using lme and MASS::glmmPQL, which is much less efficient then the penalized least squares adopted by gam and bam.
mgcv::gam– borgs