REML vs ML: Which one is better?

[dataandcrowd]

Desciption on thesis

The most popular options in GAM are REML (restricted/residual estimated maximum likelihood) and ML (maximum likelihood). The difference between the two methods is that REML is likely to be biased for the fixed effects but unbiased for the random effects, while ML is likely to be unbiased for the fixed effects but biased for the random effects. That is, REML estimates will have different mean values depending of the number of parameters but have less variances, whilst ML is vice versa. However, the difference becomes negligible when the sample size is large.

Example

• REML
  

• ML