"worrying about df (ml vs reml) is just a silly obsession of statisticians (of which I'm one)"
I too have often wondered about the importance of such tertiary issues. My half-baked understanding is that the main "practical" difference between ML vs REML is with regards to ease of computing the estimates, i.e. the REML estimates can be computed much more easily than ML. However, I am wide open to be enlightened about other practically important differences. Best, Ravi. ____________________________________________________________________ Ravi Varadhan, Ph.D. Assistant Professor, Division of Geriatric Medicine and Gerontology School of Medicine Johns Hopkins University Ph. (410) 502-2619 email: [email protected] ----- Original Message ----- From: Bert Gunter <[email protected]> Date: Tuesday, December 8, 2009 11:38 am Subject: Re: [R] lm: RME vs. ML To: 'John Sorkin' <[email protected]>, [email protected], [email protected] Cc: [email protected] > A contrarian point of view: > > If you have so little data (relative to the number of parameters to be > estimated, especially NONLINEAR parameters like covariance estimates)that > the ml vs reml bias could be large, then there's so little information > anyway that such bias is the least of your problems (identifiability > probably is a major issue-- mis-shapen confidence regions). > > Ergo, worrying about df (ml vs reml) is just a silly obsession of > statisticians (of which I'm one). > > Criticisms, public or private, welcome of course. > > This is my view only and should not be considered a stain on my > employer -- > other than its misfortune in employing me. > > > Bert Gunter > Genentech Nonclinical Biostatistics > > > -----Original Message----- > From: [email protected] [ On > Behalf Of John Sorkin > Sent: Tuesday, December 08, 2009 7:12 AM > To: [email protected]; [email protected] > Cc: [email protected] > Subject: Re: [R] lm: RME vs. ML > > Your question is well taken. I did not give any criteria because I realized > there might be different answers based upon different criteria. Certainly > one fundamental criteria would be that the estimates are BLUE, but > this is > not the only criteria one might be used. > John > -----Original Message----- > From: <[email protected]> > To: John Sorkin <[email protected]> > Cc: <[email protected]> > To: <[email protected]> > > Sent: 12/8/2009 9:39:28 AM > Subject: Re: [R] lm: RME vs. ML > > You need to give your criteria for "preferable". For normal-linear > models, REML estimates of variances are unbiased, whereas ML estimates > are > downwardly biased. My intuition is that the ML-induced bias would be > > worse in small samples. I don't know about other distributions. > Likewise I > don't know about MSE or other criterion for preference. > > > > > > > "John Sorkin" <[email protected]> > Sent by: [email protected] > 12/07/2009 09:24 PM > > To > <[email protected]> > cc > > Subject > [R] lm: RME vs. ML > > > > > > > windows XP > R 2.10 > > As pointed out by Prof. Venables and Ripley (MASS 4th edition, p275), > the > results obtained from lme using method="ML" and method="REML" are > often > different, especially for small datasets. Is there any way to > determine > which method is preferable for a given set of data? > Thanks, > john > > > John David Sorkin M.D., Ph.D. > Chief, Biostatistics and Informatics > University of Maryland School of Medicine Division of Gerontology > Baltimore VA Medical Center > 10 North Greene Street > GRECC (BT/18/GR) > Baltimore, MD 21201-1524 > (Phone) 410-605-7119 > (Fax) 410-605-7913 (Please call phone number above prior to faxing) > > Confidentiality Statement: > This email message, including any attachments, is for th...{{dropped:8}} > > ______________________________________________ > [email protected] mailing list > > PLEASE do read the posting guide > and provide commented, minimal, self-contained, reproducible code. ______________________________________________ [email protected] mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
