R help - Jan 2006 - lme model specification

I have been asked to analyse the results of (what is to me) a very
complicated experiment.

The dependent measure is the estimated distance, which is measured as a
function of the actual distance. There are also several other IVs.

The plot of log estimated distance as a function of log distance is
linear. So in the rest of the analysis I will use logestimate and
logdistance.

My plan is to see how the other IVs affect the slope and intercept of
this linear relationship between logestimate and log distance.

What complicates everything is that each datum point is not independent.
Rather, many data points come from each subject.

So:
* Each subject gets many objects at many distances which he has to
estimate.
* Each subject repeats this experiment using 4 colours of LEDs.
* Each subject repeats this experiment on 4 different sessions.
* Half the subjects do this under starlight, half under moonlight.
* Half the subjects do it with feedback and half without.

So some of these variables are within subjects and some between. I think
lme is a good way to proceed. But I am hung up on how to specify the
model

fit<-lme(fixed=logestimate~logdistance*session*illum*feedback,
random=???|subject???, data=df1)

I am familiar with the steps of model building using lm(), exploring
different models etc, so I think I will be OK once I get the idea of
specifying the basic lme model.

I have Pinheiro and Bates (2000) here.

Thanks very much for any help

Bill Simpson

Does each subject get only one LED per session or all 4 LEDs?  This 
should be important regarding which models are estimaable.  In either 
case, might the following help you?

nSubj <- 8
nSess <- 4
nObsPerSess <- 3

library(nlme)
library(e1071)
P4 <- permutations(4)

LED <- letters[t(P4[permSubj,])]

set.seed(1)
permSubj <- sample(24, nSubj)
N <- nSubj*nSess*nObsPerSess
DF <- data.frame(
   Subject=rep(1:nSubj, each=nSess*nObsPerSess),
   illum=rep(c("star", "moon"), each=N/2),
   feedback=rep(c("yes", "no"), each=N/4, length=N),
   session=rep(1:nSess, each=nObsPerSess, nSubj),
   LED=rep(LED, each=nObsPerSess),
   Rep=rep(1:nObsPerSess, nSess*nSubj),
   logdistance=rep(1:nObsPerSess, nSess*nSubj),
   logestimate=rnorm(nSubj*nSess*nObsPerSess) )

fit <- lme(logestimate~logdistance*illum*feedback+LED,
   random=~1|Subject,
   correlation=corAR1(form=~Rep|Subject/session),
   data=DF)

	  spencer graves

Bill Simpson wrote:
> I have been asked to analyse the results of (what is to me) a very
> complicated experiment.
> 
> The dependent measure is the estimated distance, which is measured as a
> function of the actual distance. There are also several other IVs.
> 
> The plot of log estimated distance as a function of log distance is
> linear. So in the rest of the analysis I will use logestimate and
> logdistance.
> 
> My plan is to see how the other IVs affect the slope and intercept of
> this linear relationship between logestimate and log distance.
> 
> What complicates everything is that each datum point is not independent.
> Rather, many data points come from each subject.
> 
> So:
> * Each subject gets many objects at many distances which he has to
> estimate.
> * Each subject repeats this experiment using 4 colours of LEDs.
> * Each subject repeats this experiment on 4 different sessions.
> * Half the subjects do this under starlight, half under moonlight.
> * Half the subjects do it with feedback and half without.
> 
> So some of these variables are within subjects and some between. I think
> lme is a good way to proceed. But I am hung up on how to specify the
> model
> 
> fit<-lme(fixed=logestimate~logdistance*session*illum*feedback,
> random=???|subject???, data=df1)
> 
> I am familiar with the steps of model building using lm(), exploring
> different models etc, so I think I will be OK once I get the idea of
> specifying the basic lme model.
> 
> I have Pinheiro and Bates (2000) here.
> 
> Thanks very much for any help
> 
> Bill Simpson
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html

If you are now presenting subjects with pairs, have you considered 
multidimensional scaling and possibly "Bradley-Terry" models & 
extensions?  RSiteSearch("multidimensional scaling") and 
RSiteSearch("Bradley-Terry") both seemed to contain potentially useful
information.

	  best wishes,
	  spencer graves

#################
Bill Simpson wrote:
 > Thanks very much Spencer for your helpful reply.
 >
 > On Fri, 2006-01-20 at 07:38 -0800, Spencer Graves wrote:
 >
 >>	  Does each subject get only one LED per session or all 4 LEDs?
 >
 > I simplified a bit. Each subject gets pairs of LEDs and needs to
 > estimate the distance between them. LED1 can be Red or Blue, and LED2
 > can be R or B -- so there are 4 combos. These are presented in random
 > order. So in one session yes each subject sees all combos of LEDs.
 >
 >
 >>  This
 >>should be important regarding which models are estimaable.  In either
 >>case, might the following help you?
 >>
 >>nSubj <- 8
 >>nSess <- 4
 >>nObsPerSess <- 3
 >>
 >>library(nlme)
 >>library(e1071)
 >>P4 <- permutations(4)
 >>
 >>LED <- letters[t(P4[permSubj,])]
 >>
 >>set.seed(1)
 >>permSubj <- sample(24, nSubj)
 >>N <- nSubj*nSess*nObsPerSess
 >>DF <- data.frame(
 >>   Subject=rep(1:nSubj, each=nSess*nObsPerSess),
 >>   illum=rep(c("star", "moon"), each=N/2),
 >>   feedback=rep(c("yes", "no"), each=N/4,
length=N),
 >>   session=rep(1:nSess, each=nObsPerSess, nSubj),
 >>   LED=rep(LED, each=nObsPerSess),
 >>   Rep=rep(1:nObsPerSess, nSess*nSubj),
 >>   logdistance=rep(1:nObsPerSess, nSess*nSubj),
 >>   logestimate=rnorm(nSubj*nSess*nObsPerSess) )
 >>
 >>fit <- lme(logestimate~logdistance*illum*feedback+LED,
 >>   random=~1|Subject,
 >>   correlation=corAR1(form=~Rep|Subject/session),
 >>   data=DF)
 >
 > Thanks very much, I didn't know about the corAR1 statement.
 >
 > Best wishes
 > Bill