[R-sig-ME] Huge speed performance difference when using non-trivial fixed effects in NLMER vs NLME
rb@rr@nt @end|ng |rom uvm@edu
Wed Aug 12 16:56:04 CEST 2020
Following Ben Bolker's methodology (described here
https://rpubs.com/bbolker/3423) I incorporated non-trivial fixed effects
in NLMER for a four-parameter logistic. I placed a reproducible
example here: https://rpubs.com/ramirob/648103
To summarize the question, if we have a dataset with individuals in
groups where we have group-specific fixed effects, NLME's performance
remains the same:
##  "NLME Time Required for data2Groups: 0.0458040237426758"
fit3Groups <- fitNLME(data3Groups,initialValues3Groups)
##  "NLME Time Required for data3Groups: 0.0375699996948242"
fit4Groups <- fitNLME(data4Groups,initialValues4Groups)
##  "NLME Time Required for data4Groups: 0.0526559352874756"
fit5Groups <- fitNLME(data5Groups,initialValues5Groups)
##  "NLME Time Required for data5Groups: 0.0502560138702393"
But when we do the analogous thing in NLMER, the performance increases
with increasing number of groups:
##  "Time required for the data2Groups: 0.404773950576782"
fitNlmer3Groups <- fitNlmer(data3Groups, initialValues3Groups)
##  "Time required for the data3Groups: 0.579570055007935"
fitNlmer4Groups <- fitNlmer(data4Groups, initialValues4Groups)
##  "Time required for the data4Groups: 0.957509994506836"
fitNlmer5Groups <- fitNlmer(data5Groups, initialValues5Groups)
##  "Time required for the data5Groups: 1.68412184715271"
In addition, NLMER is much slower in general. This is just a short
example, but for more complicated cases the differences in performance
are huge (minutes vs seconds).
Is NLMER "worth the wait" (e.g. less fragile, better convergence, etc)
when trying to do non-trivial fixed effects? Is there a better
methodology than the one described by Ben Bolker back in 2013?
Any insight appreciated. Again, you can see a reproducible example here
Ramiro Barrantes Reynolds, Ph.D.
Bioinformatics Research Associate, Microbiology & Molecular Genetics
Vermont Integrative Genomics Resource
University of Vermont
Burlington, VT 05405
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models