save also formula with saveoptsum()

[kliegl]

Is there an argument against saving the model formula along with the fitted model object?

[palday]

There would be some annoying extra bookkeeping to preserve backwards compatibility models saved without it. That's not insurmountable, but the formula doesn't preserve all the information about the model -- we don't save any of the model matrices nor the contrast coding, so that still have to be recovered elsewhere. My question in return is: what's the argument for saving the model formula?

[kliegl]

I see.

I have / had the very frequent problem that an inadvertent slight change of the model formula in the script does not allow me to restore a model object; usually a consequence of my top-down style of model selection. It is probably also interference from the save-load lme4 past. This is annoying because I usually only save model objects that take a long time to fit. No worry: I have trained myself (better: the program trained me) to reduce this problem substantially, but I am sure my frustration is shared with others ;)

Curious: Is there a best practice recommendation from your side on this? My current "style" is that, after the fit, I immediately insert the restore lines, then I comment out the fit() and saveoptsum() lines and save the script as follows:

f = @formula ...;
#m = fit(MixedModel, f, data; contrasts)
#issingular(m)
#MixedModels.PCA(m)
#saveoptsum("fits/m_optsum.json", m);

m = LinearMixedModel(f, data; contrasts);                    
restoreoptsum!(m, "./fits/m_optsum.json");

In the past, my restore lines were at the end of the script, that is far away from the formula definition. You also cannot generate the formula on the fly.

A related question: Would it be possible to have a saveoptsum() mode kwarg to generate read-only model json files?