Random slope/intercept without correlation in lmer

All you need to do is set up your own dummy variable (e.g.
ntreat=as.numeric(treat)-1, or ntreat=as.numeric(treat=="1"), or
ntreat=(original treat variable before using factor(treat) and then use
(ntreat||group) or (1|group)+(0+ntreat|group)

  This is related but not identical to the last example in ?lmer ; it's
caused by an interaction between the way that R constructs model matrices
from factors and the way lme4 uses that functionality.



On Mon, Aug 11, 2014 at 2:38 PM, Gustaf Granath <gustaf.granath at gmail.com>
wrote:

Hi
I want to model random slope and intercept without a correlation between
the two. Is it possible to do this in lmer when the predictor is a factor?

For example, imagine that x has 2 levels (control and treatment). In nlme,
I have been modeling uncorrelated intercept and slope like this:
lme(y ~ x, random=list(x = pdDiag(~ group)) ) where group is a random
factor.
It gives me the random intercept and random slope (i.e. variation in
treatment effect among groups).

In lmer, I think the corresponding model is defined as:
lmer( y ~ x + (x||rand) and I guess this gives me differences (variation
in differences to the intercept), but it includes a covariance term.

Is it possible to reproduce the above lme() model in lmer?

I have a strong feeling that Im missing something here. Most of the
literature on this subject (and R-list questions) deals with continuous
variables so pls let me know if there is a good source on this topic.

Below follows a small example.

Cheers

Gustaf


set.seed(1)
treat = rep(c(0, 1), each = 5, 10)
group = rep(1:10, each = 10)
rand.int = rep( rnorm( 10, 0, 1), each = 10)
rand.slop = rep( rnorm(10, 0, 1), each = 10)
e = rnorm(100, 0, 0.5)
y = 10 + rand.int + treat + rand.slop*treat + e
treat = factor(treat)

#lmer
library(lme4)
# with correlation between intercept and slope
mod = lmer(y ~ treat + (treat|group) )
# without correlation between intercept and slope
# gives lots of error msgs
mod2 = lmer(y ~ treat + (treat||group) )
summary(mod)
summary(mod2)
# var-covar matrix
VarCorr(mod)$group
VarCorr(mod2)$group.1 #still a covariance term

#nlme
# without correlation
library(nlme)
lme.mod <- lme(y ~ treat, random=list(group = pdDiag(~ treat)) )
summary(lme.mod)
getVarCov(lme.mod)

--
Gustaf Granath (PhD)
Post doc
McMaster University
School of Geography & Earth Sciences

	[[alternative HTML version deleted]]