[R-sig-ME] Decreasing size of gamm4 model output (multiple GB when saved to .rds)

[David Duffy]

________________________________________
From: R-sig-mixed-models <r-sig-mixed-models-bounces using r-project.org> on behalf of David Winsemius <dwinsemius using comcast.net>
Sent: Tuesday, 19 September 2023 12:24 PM
To: Meaghan Rupprecht
Cc: R-SIG-Mixed-Models using r-project.org
Subject: Re: [R-sig-ME]  Decreasing size of gamm4 model output (multiple GB when saved to .rds)

> I was puzzled by the fact that you save the output as an rds file. I suspect but am unable to confirm, that your result(s) include some representation of the data. 
> You might instead return the models with the data omitted
> David.

>> On Sep 17, 2023, at 10:11 PM, Meaghan Rupprecht <rupprecht using unbc.ca> wrote:
>>
>> I am running a Generalized Additive Mixed Model with the R package, gamm4. Each model output includes a mer object and a gam object. 
>> I need to compare 26 model structures based on a combination of variables and then conduct model averaging on the best models based on AIC.
>> The problem is, each .rds file with model output is approximately 5GB. For all 26 models, there is no way I can load them into my R environment 
>> for any kind of model comparison or averaging. Similar models run with mgcv::gamm() were only 5.7 MB, which is a much more manageable size for comparisons. 
>> (Note: unfortunately I have to stay within the gamm4 package based on model structure, otherwise I would use the mgcv model outputs).
>> Is there a way to decrease the size of the .rds file to a more reasonable size? Are there arguments within gamm4 or lme4 that could reduce the amount of 
>> extra information retained within the model?

I presume this is because gamm4's gam contains "model", which a copy of the data frame. Having little memory on my machine, I was
save()'ing then removing each gamm4 model, while tabulating summaries for each model by hand (can run deviance(), AIC() etc for each). 
Do you have the same random effects fitted in each model? I have no idea if these models can be compared if the cross-validation selected gams
have different edf's with different REs in the mer part. Maybe use gam to choose roughly equivalent polynomial fixed effects, then just do in lmer?

hth, David Duffy