[R-sig-ME] weights and lmer2

[Douglas Bates]

On 4/12/07, Sundar Dorai-Raj <sundar.dorai-raj at pdf.com> wrote:
>
> Douglas Bates said the following on 4/12/2007 11:35 AM:
> > On 4/12/07, Sundar Dorai-Raj <sundar.dorai-raj at pdf.com> wrote:
> >> Hi, all,
> >
> >> Is the weights argument supposed to work with lmer2? I'm trying to
> >> replicate some analysis from nlme::lme but am having difficulty with the
> >> weights argument. Here's some funny data to work with:
> >
> >> set.seed(42)
> >> z <- expand.grid(A = 1:5, B = 1:4, r = 1:2)
> >> n <- nrow(z)
> >> z$w <- rpois(n, 100)
> >> z$y <- rnorm(n, ifelse(z$A%%2, 1, 2) + ifelse(z$A%%2 & z$B%%2, -1, 1))
> >> z[1:2] <- lapply(z[1:2], factor)
> >
> >> ## Session 1 with lmer2
> >> library(lme4)
> >> fit <- lmer2(y ~ (1 | A) + (1 | A:B), z)
> >> wfit <- lmer2(y ~ (1 | A) + (1 | A:B), z, weights = w)
> >
> > It is supposed to work but I'm not sure if it was working in the last
> > released version of the lme4 package. I don't plan on releasing a new
> > version of lme4 until after R-2.5.0 is released.  There as changes in
> > the development version of lme4 that require R >= 2.5.0
> >
> > I'll run your test in the development version later today if someone
> > else doesn't get a chance to do so before me.  I'm tied up with
> > classes and midterm exams right now.
> >
> >> ## Session 2 with lme
> >> library(nlme)
> >> fit <- lme(y ~ 1, z, ~1 | A/B)
> >> wfit <- lme(y ~ 1, z, ~1 | A/B, weights = ~w)
> >> vfit <- lme(y ~ 1, z, ~1 | A/B, correlation = varFixed(~w))
> >>
> >> ## Standard deviations from Session 1 (lmer2)
> >>                 fit      wfit
> >> A:B      0.9371798 0.9371798
> >> A        0.8673610 0.8673610
> >> Residual 0.7908478 0.7908478
> >>
> >> ## Standard deviations from Session 2 (lme)
> >>                               fit       wfit       vfit
> >> (Intercept) | B %in% A 0.9371788 0.90186678 0.90186678
> >> (Intercept) | A        0.8673608 0.87420501 0.87420501
> >> Residual               0.7908482 0.07878061 0.07878061
> >>
> >> It appears that the weights argument has no effect for lmer2 (or for
> >> lmer, which I also tried).

>
> Hi, Prof. Bates,
>
> Thanks for the reply. I can wait for next release. I'll just continue
> using lme for now. I also forgot to add my sessionInfo if that may be
> useful:
>
>  > sessionInfo()
> R version 2.5.0 beta (2007-04-10 r41119)
> i386-pc-mingw32
>
> locale:
> LC_COLLATE=English_United States.1252;LC_CTYPE=English_United
> States.1252;LC_MONETARY=English_United
> States.1252;LC_NUMERIC=C;LC_TIME=English_United States.1252
>
> attached base packages:
> [1] "stats"     "graphics"  "grDevices" "utils"     "datasets"  "methods"
> [7] "base"
>
> other attached packages:
>         lme4      Matrix     lattice
> "0.9975-13" "0.9975-11"    "0.15-3"

Well the good news is that in the development version of lme4 the
weights argument is used.  The bad news is that it may not be used on
the proper scale.  I believe that it is being used as if it were on
the scale of the inverse of the standard deviation and the actual
usage should be on the scale of the inverse of the variance.  I'll
need the check the documentation versus the code.