Currently, I am using mixed models using “lme4” package to analyze the repeated measure data on leaf litter decomposition (using lmer function). I went through many literature but still I am not fully convinced with the use of REML = FALSE or TRUE. However, I came to a conclusion that (but not sure why different rules for fixed and random effects):
1) REML = FALSE is used in case of comparing models with different “Fixed effects” (during the simplification of model)
2) REML = TRUE is used in case of different random effects on the comparing models.
I would appreciate for the explanation on the correct use of these functions.
You may want to check this out, pretty good explanation of basics of what REML is and when to use it: http://users.stat.umn.edu/~gary/classes/5303/handouts/REML.pdf.
Tldr version: REML is a method for estimating variance components in models with random effects. If all effects are fixed, then using REML makes no sense because the first thing REML does, computationally speaking, is removing all fixed effects and evaluating remaining variance that belongs to random effects. Won't work very well if there are no random effects to be evaluated.
You may want to check this out, pretty good explanation of basics of what REML is and when to use it: http://users.stat.umn.edu/~gary/classes/5303/handouts/REML.pdf.
Tldr version: REML is a method for estimating variance components in models with random effects. If all effects are fixed, then using REML makes no sense because the first thing REML does, computationally speaking, is removing all fixed effects and evaluating remaining variance that belongs to random effects. Won't work very well if there are no random effects to be evaluated.
I suggest also this reference: Bolker BM, Brooks ME, Clark CJ, Geange SW, Poulsen JR, Stevens MHH, White J-SS, Generalized linear mixed models: a practical guide for ecology and evolution. Trends in Ecology and Evolution 24.
Best regards
Also, you don't want to use REML if you are comparing models with different fixed effects. You need to use ordinary likelihood for such cases.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series