When should I specify REML = FALSE in lmer()?

2
votes

I am running a basic Mixed Effect Model with lmer(), in R. Let say I have 2 within-subject conditions. In each condition the subject provides one measure. 

lmer(measure ~ condition + (1|subject),
      REML = TRUE, data = My_data)

The argument REML is TRUE by default. Yet, in several examples I read, people set it as FALSE.

According to the documentations "logical scalar - Should the estimates be chosen to optimize the REML criterion (as opposed to the log-likelihood)?"

Are there real differences in the estimates of fixed or random effects when I use one method or another. When should REML be TRUE? When should it be FALSE?

1
rlme4mixed-models
You shouldn't use REML fits if you want to compare two models with different fixed effects. But lme4 will refit the models in such a case for you. – Roland
Suppose I am not comparing two models. When should I specify TRUE or FALSE? – Rtist
Usually, you should use the REML fit. – Roland
The only time to use ML is when you want to compare two nested models. Otherwise, use REML fits. Why it is preferred computationally is explained by D. Bates here --> [link] (cran.r-project.org/web/packages/lme4/vignettes/Theory.pdf) The pdf is under Vignettes for lme4 package on CRAN. – mike
This question has a good answer at stats.stackexchange.com/a/272654/53514. – cbrnr

1 Answers

1
votes

Based on the comment above, a good answer can be found here:

https://stats.stackexchange.com/questions/272633/how-to-decide-whether-to-set-reml-to-true-or-false/272654#272654

To cite the author: "It’s generally good to use REML, if it is available, when you are interested in the magnitude of the random effects variances, but never when you are comparing models with different fixed effects via hypothesis tests or information-theoretic criteria such as AIC."