0

I am using the nlme package in R (4.1.1) to fit a hierarchical growth model.

The data are longitudinal (therapy sessions on level 1) nested within patients (level 2).

The data are called as follows:

value = the session score I'm predicting

EMA_Nr = identifier variable for each patient

variable = time variable

I used the gls function to fit the intercept only model:

intercept<-gls(value ~ 1, 
               data = HSCL_SG_long, 
               method = "ML", 
               na.action = na.exclude)

Then I fitted a random intercept model:

randomintercept<-lme(value ~ 1, 
                 data = HSCL_SG_long, 
                 random = ~1|EMA_Nr, 
                 method = "ML",
                 na.action = na.exclude, 
                 control = list(opt="optim"))

Then I used the update function to keep everything the same and introduce time as a predictor:

timeRI<-update(randomintercept, .~. + variable)

No problems up until here. Now the next logical step would be to introduce random slopes into the model, so the effect of time can vary from person to person:

timeRS<-update(timeRI, random = ~variable|EMA_Nr)

This is where things go south: R is computing for an eternity (I let it work 40mins). Then when I try to abort the computing, it crashes. There is no error report or anything, it just won't return any model estimates. Can anyone help me with that?

Ben Heuser
  • 1
  • 1
  • Sounds like the optimization algorithm has a hard time fitting your data to the model. How many observations do you have per person? Have you tried other optimization methods for optim, e.g. lmeControl(optimMethod = "L-BFGS-B")? You have more different optimizers available if you switch to the lme4 package. It is easy to specify exactly the same model in lme4 syntax. Your final model would look like this: `lmer(value ~ variable + (variable | EMA_Nr), ...)` – benimwolfspelz Feb 03 '22 at 10:11
  • Thank you very much for your reply! I tried other optimizers in nlme, which did not work. Same thing for the lme4 package, it crashes on the exact same line, which is the introduction of random slopes into the model. I have 15 observations per person, and interestingly it all worked smoothly when I cut it down to 5 observations per person. Will try it again with another program (SPSS) and if that won't work, I'll just use the reduced dataset. – Ben Heuser Mar 02 '22 at 10:13

0 Answers0