Thomas Lumley • 12/12/2023

Why not REML?

The article discusses the technical rationale for the svylme package not implementing REML estimation. It explains that REML corrects bias in variance components when fixed effects consume degrees of freedom, which is crucial in highly structured experimental designs. However, in survey sampling contexts where data comes from random subsamples, such designs are rare, and REML's benefits may not apply or could even lead to identifiability issues. The author concludes that without clear use cases combining complex experimental design with subsampling, adding REML is not currently planned.

0 comments

#statistics #Mixed Models #Svylme