PROC MIXED for non-normal data
in [SAS]
|
From: Christoff on 3 Feb 2010 02:22 Hello all, Can one use PROC MIXED on non-normally distributed data? I have heard it is robust to the assumptions. If so, are there any references in literature that support this? My dataset consist of body temperatures measured hourly across +-10 sequential days during summer, autumn, winter and spring. I used different study subjects (lizards) during each season, and experimental day therefore is the repeated measure. Both the number of experimental days and the number of lizards used vary among seasons resulting in an unbalanced design. PROC MIXED is the only model I know of that can handle unbalanced repeated measures data. Does anyone know of non-parametric alternatives? I have tried various transformations yet could not improve normality. Kind regards Christoff
From: Robin R High on 3 Feb 2010 09:51 Christoff, Every dataset has its own issues to work around, but first want to make sure you are basing your comments about non-normality based on a residual analysis (such as described in Chapter 10 of "SAS for Mixed Models", 2nd ed.) and not on how the original data look. GLIMMIX has some distribution alternatives that might make an improvement over the normal without computing a transformation,which works much like PROC MIXED. Robin High UNMC From: Christoff <14353075(a)SUN.AC.ZA> To: SAS-L(a)LISTSERV.UGA.EDU Date: 02/03/2010 08:29 AM Subject: PROC MIXED for non-normal data Sent by: "SAS(r) Discussion" <SAS-L(a)LISTSERV.UGA.EDU> Hello all, Can one use PROC MIXED on non-normally distributed data? I have heard it is robust to the assumptions. If so, are there any references in literature that support this? My dataset consist of body temperatures measured hourly across +-10 sequential days during summer, autumn, winter and spring. I used different study subjects (lizards) during each season, and experimental day therefore is the repeated measure. Both the number of experimental days and the number of lizards used vary among seasons resulting in an unbalanced design. PROC MIXED is the only model I know of that can handle unbalanced repeated measures data. Does anyone know of non-parametric alternatives? I have tried various transformations yet could not improve normality. Kind regards Christoff
From: Christoff on 3 Feb 2010 10:58 On Feb 3, 4:51 pm, rh...(a)UNMC.EDU (Robin R High) wrote: > Christoff, > > Every dataset has its own issues to work around, but first want to make > sure you are basing your comments about non-normality based on a residual > analysis (such as described in Chapter 10 of "SAS for Mixed Models", 2nd > ed.) and not on how the original data look. GLIMMIX has some > distribution alternatives that might make an improvement over the normal > without computing a transformation,which works much like PROC MIXED. > > Robin High > UNMC > > From: > Christoff <14353...(a)SUN.AC.ZA> > To: > SA...(a)LISTSERV.UGA.EDU > Date: > 02/03/2010 08:29 AM > Subject: > PROC MIXED for non-normal data > Sent by: > "SAS(r) Discussion" <SA...(a)LISTSERV.UGA.EDU> > > Hello all, > > Can one use PROC MIXED on non-normally distributed data? I have heard > it is robust to the assumptions. If so, are there any references in > literature that support this? > > My dataset consist of body temperatures measured hourly across +-10 > sequential days during summer, autumn, winter and spring. I used > different study subjects (lizards) during each season, and > experimental day therefore is the repeated measure. > > Both the number of experimental days and the number of lizards used > vary among seasons resulting in an unbalanced design. > PROC MIXED is the only model I know of that can handle unbalanced > repeated measures data. Does anyone know of non-parametric > alternatives? > I have tried various transformations yet could not improve normality. > Kind regards > Christoff Hi Robin Yes indeed the residual distributions are non-normal in most of the datasets. It seems as though Proc GLIMMIX might do the trick, just a quick question..I have quickly had a look at PROC GLIMMIX and noticed it has no repeated statement. Does one simply include the repeated measure in the RANDOM statement? Thank you Christoff
From: Robin R High on 3 Feb 2010 11:47 Yes the RANDOM in GLIMMIX looks much like the REPEATED in MIXED, though you need to specify residual or Rside, something like in MIXED REPEATED time / subject=id type=ar(1) R Rcorr; in GLIMMIX becomes: RANDOM time / subject=id type=ar(1) v vcorr residual; * or add Rside; Robin High UNMC From: Christoff <14353075(a)SUN.AC.ZA> To: SAS-L(a)LISTSERV.UGA.EDU Date: 02/03/2010 10:27 AM Subject: Re: PROC MIXED for non-normal data Sent by: "SAS(r) Discussion" <SAS-L(a)LISTSERV.UGA.EDU> On Feb 3, 4:51 pm, rh...(a)UNMC.EDU (Robin R High) wrote: > Christoff, > > Every dataset has its own issues to work around, but first want to make > sure you are basing your comments about non-normality based on a residual > analysis (such as described in Chapter 10 of "SAS for Mixed Models", 2nd > ed.) and not on how the original data look. GLIMMIX has some > distribution alternatives that might make an improvement over the normal > without computing a transformation,which works much like PROC MIXED. > > Robin High > UNMC > > From: > Christoff <14353...(a)SUN.AC.ZA> > To: > SA...(a)LISTSERV.UGA.EDU > Date: > 02/03/2010 08:29 AM > Subject: > PROC MIXED for non-normal data > Sent by: > "SAS(r) Discussion" <SA...(a)LISTSERV.UGA.EDU> > > Hello all, > > Can one use PROC MIXED on non-normally distributed data? I have heard > it is robust to the assumptions. If so, are there any references in > literature that support this? > > My dataset consist of body temperatures measured hourly across +-10 > sequential days during summer, autumn, winter and spring. I used > different study subjects (lizards) during each season, and > experimental day therefore is the repeated measure. > > Both the number of experimental days and the number of lizards used > vary among seasons resulting in an unbalanced design. > PROC MIXED is the only model I know of that can handle unbalanced > repeated measures data. Does anyone know of non-parametric > alternatives? > I have tried various transformations yet could not improve normality. > Kind regards > Christoff Hi Robin Yes indeed the residual distributions are non-normal in most of the datasets. It seems as though Proc GLIMMIX might do the trick, just a quick question..I have quickly had a look at PROC GLIMMIX and noticed it has no repeated statement. Does one simply include the repeated measure in the RANDOM statement? Thank you Christoff
From: Dale McLerran on 3 Feb 2010 13:25 It is correct that in the GLIMMIX procedure the residual covariance structures are specified through a RANDOM statement instead of a REPEATED statement. In trying to match a PROC MIXED REPEATED statement, there are two variants of the GLIMMIX RANDOM statement that are employed, depending on whether or not your REPEATED statement would or would not specify a time/space effect. That is, the PROC MIXED REPEATED statement could be constructed as: repeated time / subject=subjID type=AR(1); or as repeated / subject=subjID type=AR(1); Note that I find the latter to be dangerous for most covariance specifications. It is fine for TYPE=CS, but for an AR(1) specification, I always prefer to use an explicit specification of the repeated measure effect (TIME in the example above). Anyway, given these two different constructions of a REPEATED statement which one might employ for PROC MIXED, there are corresponding RANDOM statements to use in the GLIMMIX procedure as follows: random time / subject=subjID type=AR(1) residual; or random _residual_ / subject=subjID type=AR(1); HTH, Dale --------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra(a)NO_SPAMfhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 --------------------------------------- --- On Wed, 2/3/10, Robin R High <rhigh(a)UNMC.EDU> wrote: > From: Robin R High <rhigh(a)UNMC.EDU> > Subject: Re: PROC MIXED for non-normal data > To: SAS-L(a)LISTSERV.UGA.EDU > Date: Wednesday, February 3, 2010, 8:47 AM > Yes the RANDOM in GLIMMIX looks much > like the REPEATED in MIXED, though > you need to specify residual or Rside, something like > > in MIXED > > REPEATED time / subject=id type=ar(1) R Rcorr; > > in GLIMMIX becomes: > > RANDOM time / subject=id type=ar(1) v vcorr residual; * or add Rside; > > > Robin High > UNMC > > > > > > > From: > Christoff <14353075(a)SUN.AC.ZA> > To: > SAS-L(a)LISTSERV.UGA.EDU > Date: > 02/03/2010 10:27 AM > Subject: > Re: PROC MIXED for non-normal data > Sent by: > "SAS(r) Discussion" <SAS-L(a)LISTSERV.UGA.EDU> > > > > On Feb 3, 4:51 pm, rh...(a)UNMC.EDU > (Robin R High) wrote: > > Christoff, > > > > Every dataset has its own issues to work around, but > first want to make > > sure you are basing your comments about non-normality > based on a > residual > > analysis (such as described in Chapter 10 of "SAS for > Mixed Models", 2nd > > ed.) and not on how the original data > look. GLIMMIX has some > > distribution alternatives that might make an > improvement over the normal > > without computing a transformation,which works much > like PROC MIXED. > > > > Robin High > > UNMC > > > > From: > > Christoff <14353...(a)SUN.AC.ZA> > > To: > > SA...(a)LISTSERV.UGA.EDU > > Date: > > 02/03/2010 08:29 AM > > Subject: > > PROC MIXED for non-normal data > > Sent by: > > "SAS(r) Discussion" <SA...(a)LISTSERV.UGA.EDU> > > > > Hello all, > > > > Can one use PROC MIXED on non-normally distributed > data? I have heard > > it is robust to the assumptions. If so, are there any > references in > > literature that support this? > > > > My dataset consist of body temperatures measured > hourly across +-10 > > sequential days during summer, autumn, winter and > spring. I used > > different study subjects (lizards) during each season, > and > > experimental day therefore is the repeated measure. > > > > Both the number of experimental days and the number of > lizards used > > vary among seasons resulting in an unbalanced design. > > PROC MIXED is the only model I know of that can handle > unbalanced > > repeated measures data. Does anyone know of > non-parametric > > alternatives? > > I have tried various transformations yet could not > improve normality. > > Kind regards > > Christoff > > Hi Robin > > Yes indeed the residual distributions are non-normal in > most of the > datasets. > It seems as though Proc GLIMMIX might do the trick, just a > quick > question..I have quickly had a look at PROC GLIMMIX and > noticed it has > no repeated statement. Does one simply include the repeated > measure in > the RANDOM statement? > > Thank you > Christoff >
|
Next
|
Last
Pages: 1 2 Prev: Making a table of % of missing values for several variables Next: lag calculation macro |