Subscribe Bookmark RSS Feed

Method to use in random effects model?

kkoprow

Community Trekker

Joined:

Jun 15, 2016

I wish to create a random effects model for some data. To do so I select the "random effect" attribute for my model effect variable. I see from this page that I can choose between REML and EMS as my model fitting method. My questions are: When would I select one over the other and is there any difference between the two? I have attached a sample data set from Montgomery's text.

1 ACCEPTED SOLUTION

Accepted Solutions
Solution

REML estimates are independent of the fixed effects in the models. When data are balanced, or nearly balanced, this will make little difference and the estimates should be close. When data are unbalanced, EMS estimates may become adversely effected (even to the point of giving you negative variance components). You don't lose anything by sticking with the default REML. About the only time I would use EMS is when I wanted to match historical results in cases where EMS was used.

Don

3 REPLIES
Solution

REML estimates are independent of the fixed effects in the models. When data are balanced, or nearly balanced, this will make little difference and the estimates should be close. When data are unbalanced, EMS estimates may become adversely effected (even to the point of giving you negative variance components). You don't lose anything by sticking with the default REML. About the only time I would use EMS is when I wanted to match historical results in cases where EMS was used.

Don

kkoprow

Community Trekker

Joined:

Jun 15, 2016

Sorry for the late response. Thank you for your answer!

jdkurtis

New Contributor

Joined:

Dec 14, 2016

As a quick follow-on.... I am trying to fit a random effect in a model that has a binary outcome. JMP13Pro does not seem to be able to do this... Is there a work-around ?

thx

jake

 

PS- i've been asking JMP tech support to implement this since JMP6.. :((