Reduce model object size

[acircleda]

I am building a production level pipeline using around 60 different models all fit with a relatively simple formula :y ~ x + z + (1 | id) , however some of the model files are massive (~2gb) because of the number of observations I use.

I need the model object to predict on new data. I am wondering what are the basic elements of the model object necessary for predict? I'd like to place them into a new, slimmer object and work from that. I have already taken a deductive approach, nullifying elements that contribute to a high object size (I got the 2gb file down to 119 mb) but wondering if I could take a better approach.

[bbolker]

Depending on what you need to do, you could make a very slim pipeline by saving only the coefficients and implementing your own predict method, i.e.

beta <- fixef(model)$cond
b <- getME(model, "b")  ## needs development version of the package
X <- model.matrix(~x + z, newdata)
Z <- model.matrix(~0+id, newdata)
pred <- X %*% beta + Z %*% b

... but of course that doesn't get you things like back-transformation via the inverse link in a GLMM, standard errors of predictions, etc etc etc ...