Beginner help setting up a mixed model

Hi Beth,

On Tue, Oct 12, 2010 at 2:03 PM, Beth Holbrook <bvholbi at yahoo.com> wrote:

I am new to both R and mixed models, so I apologize in advance about the
simplistic nature of my question.

My dataset is unbalanced and have been I have been advised by the
statistician
on my Ph.D. committee to run a mixed model. ?The response variable is
distance.
The factors are preyswim, light, and trial. ?I am interested in the effects
of
light and preyswim on distance. ?Trial is a random factor and there is also
a
crossed random effect of preyswim*trial. ?In SAS, the code would look like
this:

Proc mixed data=laketrout;
Class preyswim light trial;
Model distance = light | preyswim;
Random trial trial*preyswim;
Run;

My understanding is that to test the full ANOVA model in R, I would need to
code
three separate models and determine the most appropriate. ?This is how I
attempted to code the same model using the lme4 package in R:

options(contrasts=c(unordered="contr.SAS", ordered="contr.poly"))
mixed1 <- lmer(distance~preyswim*light + (1|trial) + (1|preyswim),
laketrout,
REML=FALSE)

I am not sure this is the model you want to run. If you want a crossed
random effect I believe you need something like the following:

mixed1 <- lmer(distance~ + (1|trial) + (1|preyswim) + (1|preyswim:light),
laketrout,REML=FALSE)

But then you no longer have any fixed effects. So I am not sure. I wonder if
you want this based on what you said above...

mixed1 <- lmer(distance~ preyswim + light + (1|trial) + (1|preyswim:light),
laketrout,REML=FALSE)

or ...

mixed1 <- lmer(distance~ preyswim + light + preyswim*light + (1|trial),
laketrout,REML=FALSE)

But that model doesn't treat preyswim*light as a crossed random effect


mixed2 <-lmer(distance~preyswim + light + (1|trial) + (1|preyswim),

laketrout,
REML=FALSE)
mixed3 <-lmer(distance~preyswim + (1|trial) + (1|preyswim), laketrout,
REML=FALSE)

My questions are:

1. ?Did I correctly set up the mixed model in R?
2. ?When determining the most appropriate model in R (I used the anova
function
to compare the three models), does it matter whether I use the AIC/BIC or
log-likelihood criterion? ?Along those lines, I'm unsure whether I need to
use
REML=TRUE or REML=FALSE in my code.

The anova function would work here because your models are nested. I believe
it's quite common to use the AIC and use the criterion of Burnham and
Anderson, 2002. Also, I believe you want REML=TRUE when comparing models
with different random effects.

I no longer have access to SAS although it is the preferred statistical
program
used by my committee statistician. ?I understand that there are some
limitations
to SAS anyway, particularly when calculating p-values for mixed models with
random effects.

My preference is to use R for my statistical analyses, as
long
as I know that I'm using it correctly.

Thanks so much for your help,
-Beth

Beth Holbrook
University of Minnesota
bvholbi at yahoo.com




? ? ? ?[[alternative HTML version deleted]]

--
Christopher David Desjardins
Ph.D. student, Quantitative Methods in Education
M.S. student, Statistics
University of Minnesota

? ? ? ?[[alternative HTML version deleted]]