Restricted Maximum Likelihood (REML) vs Maximum Likelihood (ML) Analysis
Two types of analysis choices are available for most split-plot (mixed model) designs: Maximum Likelihood (ML) and Restricted Maximum Likelihood (REML). ML differs from REML in how the random effects are estimated and p-values computed for the fixed effects. (Note that ML is not available for combined split-plot designs.)
REML is the default in most cases because it provides unbiased estimates, while ML is only unbiased for large designs. In the software, this threshold is defined by the criteria that the number of groups is greater than 500 and the number of runs is greater than 2000. For these large designs, the ML estimates should be very close to the REML estimates and the default selection will switch to ML for faster analysis.
Note
ML, not REML, is always used for automatic model selection. REML cannot be used to compare models that differ in their fixed effects because REML estimates the random effects by removing the fixed effects. This means that the statistics listed on the Model Selection Log are for ML and may differ from those listed on the Model Comparison Statistics tab when REML is selected for the mixed model.
References
Shayle R. Searle, George Casella, and Charles E. McCulloch. Variance Components. John Wiley & Sons, Inc., 1992.