scope, lme, ns, nlme, splines

Your solution works beautifully for predict.lme in the same data that
were used to compute the model. But what if I want to compute the
population fitted values in newdata?

predict.lm is able to find the object called "fixed". But predict.lme is
unable to find it and returns an error.

library("nlme")
library(splines)
library(ggplot2)
set.seed(5)
junk<-data.frame(x=rnorm(100))
junk$y<-junk$x + rnorm(nrow(junk))
junk$idvar<- factor(sample(LETTERS[1:10], size=nrow(junk), replace=TRUE))
PRETTYNEWDATA<-data.frame(x=seq(-2, 2, length=20))

fitfunction<-function(splineDF=1) {
        fixed <- eval(
           substitute(
                 expr= y ~ ns(x, df=splineDF)
                 , env= list(splineDF=splineDF)
              )
           )
  		thislm<- lm( fixed , data= junk)
  		junk$fitlm<-predict(thislm)
        thislme<-lme( fixed= fixed , random= ~ 1 | idvar , data= junk)
        junk$fitlme<-predict(thislme,level=0)
        print(
           ggplot( junk, aes(x, y))
              + geom_point(shape=1)
              + geom_line( aes(x, fitlme), color="red", lwd=2)
              + geom_line( aes(x, fitlm), color="blue", lty=3)
  				+ ggtitle(deparse(fixed))
           )
        PRETTYNEWDATA$fitlm<-predict(thislm,newdata=PRETTYNEWDATA)
  		print(head(PRETTYNEWDATA))
  	cat("about to use newdata with lme \n")
        PRETTYNEWDATA$fitlme<-predict(thislme,level=0, newdata=PRETTYNEWDATA)
        thislme
}

Jake Wegelin

On Thu, 6 Dec 2012, Duncan Murdoch wrote:

On 06/12/2012 2:28 PM, Jacob Wegelin wrote:

I want to fit a series of lme() regression models that differ only in the
degrees of freedom of a ns() spline. I want to use a wrapper function to do
this. The models will be of the form

y ~ ns(x, df=splineDF)

where splineDF is passed as an argument to a wrapper function.

This works fine if the regression function is lm(). But with lme(),
I get an error.  fitfunction() below demonstrates this.

Why?

Presumably this is a bug in nlme.  Formulas have associated environments.  In
your example, that would be the local frame of the fitfunction, and splineDF
lives there.  It looks as though lme is not looking in the formula
environment for the splineDF value.

Another workaround besides the one you found is to use substitute() to put
the value of splineDF into the formula, i.e.

fitfunction<-function(splineDF=1) {
      junk$splineDF <- splineDF
     thislm<-lm( y ~ ns(x, df=splineDF), data= junk)
     thislm
     cat("finished thislm\n")
      fixed <- eval(substitute( y ~ ns(x, df=splineDF),
list(splineDF=splineDF)))
     thislme<-lme(
         fixed= fixed
         , random= ~ 1 | idvar
         , data= junk)
     thislme
}

It's not much better than your kludge in this example, but it might be more
convenient if you want to work with different formulas.

Duncan Murdoch

KLUDGEfit() below provides a clumsy solution. It turns the lme()
command, along with the appropriate value of splineDF, into a text
string and then uses eval(parse(text=mystring)).  But is there not a
more elegant solution?

set.seed(5)
junk<-data.frame(x=rnorm(100))
junk$y<-junk$x + rnorm(nrow(junk))
junk$idvar<- factor(sample(LETTERS[1:10], size=nrow(junk), replace=TRUE))
library("nlme")
library(splines)

fitfunction<-function(splineDF=1) {
   	thislm<-lm( y ~ ns(x, df=splineDF), data= junk)
   	thislm
   	cat("finished thislm\n")
   	thislme<-lme(
   		fixed= y ~ ns(x, df=splineDF)
   		, random= ~ 1 | idvar
   		, data= junk)
   	thislme
}
fitfunction()

KLUDGEfit<-function(splineDF=2) {
   	thislm<-lm( y ~ ns(x, df=splineDF), data= junk)
   	thislm
   	cat("finished thislm\n")
   	mystring<-paste(
   	"thislme<-lme( fixed= y ~ ns(x, df="
   	, splineDF
   	, "), random= ~ 1 | idvar , data= junk)"
   	, sep="")
   	eval(parse(text=mystring))
   	thislme
}
KLUDGEfit()

Thanks for any insight

Jacob A. Wegelin
Assistant Professor
Department of Biostatistics
Virginia Commonwealth University
830 E. Main St., Seventh Floor
P. O. Box 980032
Richmond VA 23298-0032
U.S.A.