Date:         Wed, 2 Jun 2004 15:08:48 +0200
Reply-To:     gregor@mrcina.bfro.uni-lj.si
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Gregor Gorjanc <gregor@MRCINA.BFRO.UNI-LJ.SI>
Organization: University of Ljubljana
Subject:      Re: The mixed model with spatial correlated data
Comments: To: Marine Castaing <castaing@iarc.fr>
In-Reply-To:  <40BDCDEE.2080103@iarc.fr>
Content-Type: text/plain; charset="iso-8859-1"

> >>
> >>>>Right now, I have 2 questions about it:
> >>>>- Can I fit such a model given that these data are not individual data
> >>>>but data of countries?
> >
> >Do you actually have indvidual data?
> >I have incidence within countries and I did not know if mixed model was
> > adapted to such data. I do not have any individual data for which we
> > could have the place where they live for example, I hope you see what I
> > mean.

Can you type some data. Previously you typed

country1 n1 n2 n3
country2 n1 n2 n3
country3 n1 n2 n3

where n1, n2 and n3 are incidences for cancers. This means that you dont have
individual data but grouped data for each country.

About grouped and individual data. Well, it is better to use indvidual data,
since you lose variablity with grouped data.

> >
> >>>>- Can we decide to choose the ML method rather than the REML one if
> >>>>the AIC criteria is better?
> >
> >ML method differs from REML that it does not account for degress of
> > freedom used for fixed effects. I think that AIC or any other values
> > (BIC, LRT) should not be compared between ML and REML method, since
> > loglikelihood is not the same, due to mentioned fact.
>
> It is true, but in theory it is only said that these methods are the same
> for big set of data or that when one is interested in estimates of variance
> parameters it is better to use REML and above all that ML method is biased,
> always because of freedom degrees!!! but that do not tell whether it is
> better to use ML or REML when you have 38 countries pooled in 6 groups.
> Another way to do is to do with ML and compare with the REML results.
> Because I am interested in p-value of kidney parameter in my case, I
> compare and they are the same so I do not know what to tell about the
> reason I chose ML rather than REML as methodological aspect.

ML estimates might be biased. Therefore REML is prefered. If results are the
same, even better --> and again use REML.

--
Lep pozdrav / With regards

Gregor