Thanks for your help.
Rickie Domangue Dept. of Math & Stats James Madison University --On Monday, May 20, 2002 8:48 AM -0400 "E. Jacquelin Dietz" <[EMAIL PROTECTED]> wrote: > Here are comments from two of my colleagues at NC State in response to > Rickie Domangue's message about proc mixed (see below). > > Jackie Dietz > > > From Joy Smith: > ------------------------------- > There are some more examples for proc mixed on the SAS web site. > > Here is a link to the statistical samples: > http://ftp.sas.com/techsup/download/stat/ > > There also some in the Proc Mixed section of the SAS OnLineDoc. > OnlineDoc is > available on the SAS web site at: > http://v8doc.sas.com/sod_register.html > ------------------------------- > > > From Dave Dickey: > ------------------------------- > I believe the reason you do not see time*block in a random statement > is that this would imply an additive random component that changes > each time a new (block,time) level is encountered. If you treat > this as a repeated measures type of situation, then a term like > time(block*A) would seem to be sufficient to model this kind of > effect. > > The random statement sets up some matrices of (often) dummy variables > and you could think of the repeated statement in a similar vein. You > essentially want to avoid specifying two or more effects whose columns > span the same space (your dependency question). > > In fact the effect of the repeated statement can often be reproduced by > a careful invocation of a corresponding random statement. In the > program > below I generate and analyze data from a split plot with repeated > measures on the whole plots. Notice that I have used the RANDOM > statement rather than the repeated statement to get my repeated > measures analysis on the whole plots. My model is > Y(i,j,k,t) = mu + Block(i)+ A(j) + D(i,j) + w(i,j,,t) > + B(k) + AB(j,k) + e(i,j,k,t) > As it stands, the whole plot error is D(i,j)+w(i,j,t) > and I have let the w(i,j,t) have an autoregressive order 1 structure. > You could also have a model without the D, or of course fit both > and test D (I would recommend not using the default Wald tests - > do a likelihood ratio calculation) > If you follow this program more or less, you should not have > any dependency problems. > D. A. Dickey > NCSU > > ===============PROGRAM =================================== > data a; et=normal(1827655)/sqrt(1-.5*.5); > do block = 1 to 4; BEFF=8*normal(9876655); > do A = 1 to 2; > EA = 2*normal(1827655); > do time=1 to 5; > ET = .5*et + normal(1827655); > Ypart = BEFF + 2*A +EA + ET; > do B = 1 to 3; > Y=Ypart - 3*B +A*B + normal(1827655); > output; end;end; end; end; > proc glm; class a b block; > model Y = block a block*a b a*b; > proc mixed; class block a b time; > model Y = a|b; > random block a*block; > random time(a*block)/type=ar(1); > repeated time/subject=a*b*block > /*type=ar(1)*/; > * This would fit an AR(1) structure within the split plots ; > run; > ------------------------------- > >> Rickie Domangue wrote: >> >> > Can anyone help or provide resources for >> > help with the following mixed model analysis questions >> > in SAS proc mixed? >> > >> > 1. Understanding when specifications for the random >> > and repeated statements for repeated measures designs >> > result in confounding of certain variance components. >> > SAS System for Mixed Models provides a couple of examples. >> > Are there other places that I could look? In particular, >> > I'm looking for the situation when the design is a split >> > plot with whole plots blocked and repeated measures taken >> > on the whole plots. >> > >> > 2. In a split plot design, with levels of factor A assigned >> > to whole plots, with whole plots blocked, and repeated >> > measures over time on the whole plots, including or not >> > including the block x time interaction in the >> > random statement. In most examples of split plot design analysis >> > in proc mixed that I see, this term is omitted. >> > >> > Thanks for any help. >> > >> > Rickie Domangue >> > Dept. of Math & Stats >> > James Madison University >> > >> > email: [EMAIL PROTECTED] >> > phone: 540-896-4232 >> > >> > -- >> > Domangue, Rickie James >> > [EMAIL PROTECTED] > > -- > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ > E. Jacquelin Dietz (919) 515-1929 (phone) > Department of Statistics, Box 8203 (919) 515-1169 (FAX) > North Carolina State University > Raleigh, NC 27695-8203 USA [EMAIL PROTECTED] > > Street address for FedEx: > Room 210E Patterson Hall, 2501 Founders Drive > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ > . > . > ================================================================= > Instructions for joining and leaving this list, remarks about the > problem of INAPPROPRIATE MESSAGES, and archives are available at: > . http://jse.stat.ncsu.edu/ . > ================================================================= -- Domangue, Rickie James [EMAIL PROTECTED] . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
