I'm fitting a mixed effects model:
fit.1 <- lme(y~x,random=~1|id,data=df)
There are two different observations for each id for both x and y. When
I use plot(fit.1), there is a strong increasing linear trend in the
residuals versus the fitted values (with no outliers). This also happens
if I use random=~x|id. Am I specifying something incorrectly?
Rick B.
R-help (AT) stat (DOT) math.ethz.ch mailing list
PLEASE do read the posting guide
and provide commented, minimal, self-contained, reproducible code.

8/9/06, Rick Bilonick <rab45+@pitt.eduwrote:
I'm fitting a mixed effects model:

fit.1 <- lme(y~x,random=~1|id,data=df)

There are two different observations for each id for both x and y. When
I use plot(fit.1), there is a strong increasing linear trend in the
residuals versus the fitted values (with no outliers). This also happens
if I use random=~x|id. Am I specifying something incorrectly?

Could you provide a reproducible example please?

I suspect that the problem comes from having only two observations per
level of id. When you have very few observations per group the roles
of the random effect and the per-observation noise term in explaining
the variation become confounded. However, I can't check if this is
the case without looking at some data and model fits.

R-help (AT) stat (DOT) math.ethz.ch mailing list

PLEASE do read the posting guide
and provide commented, minimal, self-contained, reproducible code.

Wed, 2006-08-09 at 15:04 -0500, Douglas Bates wrote:
8/9/06, Rick Bilonick <rab45+@pitt.eduwrote:
I'm fitting a mixed effects model:

fit.1 <- lme(y~x,random=~1|id,data=df)

There are two different observations for each id for both x and y. When
I use plot(fit.1), there is a strong increasing linear trend in the
residuals versus the fitted values (with no outliers). This also happens
if I use random=~x|id. Am I specifying something incorrectly?

Could you provide a reproducible example please?

I suspect that the problem comes from having only two observations per
level of id. When you have very few observations per group the roles
of the random effect and the per-observation noise term in explaining
the variation become confounded. However, I can't check if this is
the case without looking at some data and model fits.

Unfortunately, I can't send the actual data. I did make a simple
intercepts-only example with two observations per group but it does not
exhibit the linear trend.

library(nlme)

x <- rnorm(20,5,1)
id <- factor(rep(1:20,each=2))

y <- as.vector(sapply(x,rnorm,n=2,sd=0.2))

df <- data.frame(id,y)
df.gd <- groupedData(y~x|id,data=df)

summary(lme.1 <- lme(y~1,random=~1|id,data=df.gd))
plot(lme.1)

If I fit an intercepts-only model to the actual data, I still see the
trend in the residuals.

What other analysis would you suggest?

Rick B.

R-help (AT) stat (DOT) math.ethz.ch mailing list

PLEASE do read the posting guide
and provide commented, minimal, self-contained, reproducible code.

Wed, 2006-08-09 at 15:04 -0500, Douglas Bates wrote:
8/9/06, Rick Bilonick <rab45+@pitt.eduwrote:
I'm fitting a mixed effects model:

fit.1 <- lme(y~x,random=~1|id,data=df)

There are two different observations for each id for both x and y. When
I use plot(fit.1), there is a strong increasing linear trend in the
residuals versus the fitted values (with no outliers). This also happens
if I use random=~x|id. Am I specifying something incorrectly?

Could you provide a reproducible example please?

I suspect that the problem comes from having only two observations per
level of id. When you have very few observations per group the roles
of the random effect and the per-observation noise term in explaining
the variation become confounded. However, I can't check if this is
the case without looking at some data and model fits.

I tried using geeglm from geepack to fit a marginal model. I understand
this is not the same as a mixed effects model but the residuals don't
have the linear trend. Should I avoid using lme in this case?

Rick B.

R-help (AT) stat (DOT) math.ethz.ch mailing list

PLEASE do read the posting guide
and provide commented, minimal, self-contained, reproducible code.

8/9/06, Rick Bilonick <rab45+@pitt.eduwrote:
Wed, 2006-08-09 at 15:04 -0500, Douglas Bates wrote:
8/9/06, Rick Bilonick <rab45+@pitt.eduwrote:
I'm fitting a mixed effects model:

fit.1 <- lme(y~x,random=~1|id,data=df)

There are two different observations for each id for both x and y. When
I use plot(fit.1), there is a strong increasing linear trend in the
residuals versus the fitted values (with no outliers). This also happens
if I use random=~x|id. Am I specifying something incorrectly?

Could you provide a reproducible example please?

I suspect that the problem comes from having only two observations per
level of id. When you have very few observations per group the roles
of the random effect and the per-observation noise term in explaining
the variation become confounded. However, I can't check if this is
the case without looking at some data and model fits.

I tried using geeglm from geepack to fit a marginal model. I understand
this is not the same as a mixed effects model but the residuals don't
have the linear trend. Should I avoid using lme in this case?

Probably.

R-help (AT) stat (DOT) math.ethz.ch mailing list

PLEASE do read the posting guide
and provide commented, minimal, self-contained, reproducible code.