Best practices for saving/exporting a model

[amillscampisi]

Is there a best way to save/export a model? I am estimating several different versions of a model and I would like to archive the results as I go.

[dmbates]

I hadn't considered save/restore kinds of operations.

You may want to check how large a file would be generated by @save in the JLD2 or JLD packages.

My initial thought was what would be the minimal amount of information to save so as to regenerate the fitted model and I answer that below. However, it may not be all that difficult or wasteful to save the whole fitted model.

For a LinearMixedModel the only parameters that must be saved are the θ parameters, which determine the relative covariance factor. If you save the formula and data table that generated the model then you should be able to restore the model with

updateL!(setθ!(LinearMixedModel(form, data), θ))

It would probably be a good idea to save the optsum field of the fitted model if you want to display it. There is a check in the show method on whether the model has been fit and that is determined by checking the number of function evaluations performed in the optimization.

For a GeneralizedLinearMixedModel things get a bit more complicated but you may be able to get away with saving the information to regenerate the model object (formula, data, distribution, link), the fixed effects and the θ parameters.

[amillscampisi]

In the GeneralizedLinearMixedModel case, what would the consequence be of not saving the data? I am building an application that is going to be estimating a model every day. I need to save the model to track drift over time, and to do re-runs of the model. Saving the data for each day would be prohibitive.

[Nosferican]

Depending on what you need to save, you could have a function save a subset of the model. However, depending on what you mean by re-running the model the strategy would change. For example, if it is only the data the issue, maybe in-place empty those fields in the struct.

[palday]

If you just want to archive the estimates/results from the daily model and not the model per se then you could simply save the relevant parts of the model. For example, to extract the fixed-effects, use coeftable(m); for the random-effects VarCorr(m), for the conditional modes ranef(m).

[dhalpern]

I am in a similar setting where I'd like to run a model on a cluster and then do post processing on the estimates/prediction after fitting. Is this still the recommended approach? And is there and example for what would be necessary with a GeneralizedLinearMixedModel? Thanks!