Date: Wed, 7 Oct 2009 09:51:24 0700
ReplyTo: Dale McLerran <stringplayer_2@YAHOO.COM>
Sender: "SAS(r) Discussion" <SASL@LISTSERV.UGA.EDU>
From: Dale McLerran <stringplayer_2@YAHOO.COM>
Subject: Re: Appropriate model for intraclass correlation coefficient
InReplyTo: <eb875ad3318d40e9b293c257be1e5b7d@b18g2000vbl.googlegroups.com>
ContentType: text/plain; charset=iso88591
 On Tue, 10/6/09, cat.. <cat.b41@GMAIL.COM> wrote:
> From: cat.. <cat.b41@GMAIL.COM>
> Subject: Re: Appropriate model for intraclass correlation coefficient
> To: SASL@LISTSERV.UGA.EDU
> Date: Tuesday, October 6, 2009, 10:55 PM
> On Oct 5, 7:27 pm, stringplaye...@YAHOO.COM
> (Dale McLerran) wrote:
> > Cat,
> >
> > With only a single assessment from any one rater (one
> child,
> > one parent of that child), I don't see how you can
> estimate
> > a random rater effect. Assuming there might be
> some sort
> > of generation effect (parents tend to respond
> differently
> > than the child), the model which I would use for
> anICCin
> > this setting is:
> >
> > 4  random family effect and fixed generation effect
> >
> > and estimate theICCfrom code something like the
> following:
> >
> > proc mixed data=mydata;
> > class family generation;
> > model y = generation / s;
> > random intercept / subject=family;
> > run;
> >
> > This code will give you two variance components, the
> family
> > variance component (Vf) and the residual variance
> (Vresid).
> > TheICCwould be obtained as:
> >
> > ICC= Vf / (Vf + Vresid)
> >
> > I would note that this model is a variant of your
> model 3a.
> > But if you have a fixed rater effect and every parent
> and
> > every child is a rater, then the fixed rater effects
> in
> > model 3a would fully parameterize the model even
> before you
> > include random patient (or, in my presentation,
> random
> > family) effects. Note, too, that there would be
> no error
> > df for the residual term when every response is from
> a
> > different rater. This is a problem for a
> normally
> > distributed response (which is what I presume you wish
> to
> > assume).
> >
> > Dale
> > 
> > Dale McLerran
> > Fred Hutchinson Cancer Research Center
> > mailto: dmclerra@NO_SPAMfhcrc.org
> > Ph: (206) 6672926
> > Fax:(206) 6675977
> > 
> >
> >  Show quoted text 
>
> Thanks Dale for this quick answer !
>
> I agree with the fixed effect for the generation. In fact,
> the model
> you mention is exactly y = patient (random) + rater (fixed)
> + error.
> Family effect = patient effect as we have no sibling in the
> sample.
> rater effect = generation effect.
>
> What I'm less clear about is the code. I'm not very
> familiar with proc
> mixed and I wonder how you can fit a model y = family +
> generation and
> get a Vf variance with a statement like "model y =
> generation" instead
> of "model y = family generation".
>
> Would the code
>
> proc mixed data=mydata;
> class family generation;
> model y = family generation / s;
> random family;
> run;
>
> yield the same results ?
>
> Also, what about the  required  independance between
> family and
> generation when they are in fact patient and rater and the
> rater is
> either the patient him/herself or one of his/her parents ?
>
> Catherine.
>
Catherine,
The MIXED procedure employs code which is similar to, but
not identical to, GLM procedure code.
When using the GLM procedure, you name every effect of
interest on the MODEL statement. You then declare which of
those effects are random on the RANDOM statement. Thus,
the MODEL statement for the GLM procedure names both
fixed and random effects.
When using the MIXED procedure, the MODEL and RANDOM
statements both identify effects that enter the model. The
MODEL statement identifies only those effects which enter
into the model as fixed effects while the RANDOM statement
identifies those effects which enter the model as random
effects. Thus, in the MIXED procedure, fixed and random
effects do not appear together on the MODEL statement.
If you were to name family as an effect on both the MODEL
and RANDOM statements of the MIXED procedure, you would be
declaring that FAMILY has a fixed effect and a random effect.
This is not at all what you want. In fact, this would be
an overparameterized model.
As far as your question about independence between family
and rater, I don't believe there is any such requirement.
You need to be able to assume that the random family
effects are independent of the personspecific residuals.
That is a trivial assumption. After conditioning on the
family effects, the personspecific residuals must be
independent of the family effects. I would note, though,
that since you have a balanced design with one child and
one parent completing the assessment (I am correct about
that, aren't I?), then the generation effect is totally
independent of family effects.
By the way, I don't understand your equivalence statement,
that
patient(random) + rater(fixed) + error
and
family(random) + generation(fixed) + error
represent the same model. To my way of thinking, every
individual is a different rater. There would be as many
raters as there are assessments. Put another way, in
each family, there would be two raters  parent and child.
The raters in family i would be different from the raters
in family j. Thus, the rater effect would exactly
produce the assessment.
The generation effect represents a systematic difference
between parent and child ratings. This difference is a
1 df effect, regardless of the number of families in
the data set.
Perhaps it is just terminology, that by rater you mean
parent or child. That just isn't terminology which I
would use. I think of rater as the person who completed
the assessment.
Dale

Dale McLerran
Fred Hutchinson Cancer Research Center
mailto: dmclerra@NO_SPAMfhcrc.org
Ph: (206) 6672926
Fax: (206) 6675977

