R help - Sep 2007 - R lmer with problem of 'sd slot has negative entries'

Dear R Users, 

I want to fit GLMM with lmer with binomial data and a one-way random
effects model with an overall mean and random effects. From R help,
Laplace is slower than PQL, but more accurate. When I fit my model with
Laplace method with 

 

control = list (usePQL = FALSE)), 

 

for most data sets it works well, but for some I get an error message 

 

(Error in if (any(sd < 0)) return("'sd' slot has negative
entries") :
missing value where TRUE/FALSE needed) 

 

In these cases I get an estimate for the fixed effect but do not get an
estimate for se.fixef.  If I change the method to PQL or change 

 

control = list (usePQL = TRUE) 

 

then I get estimates for both fixed effect and se.fixef, but in the
example below the estimates of fixed effects are different for 3 of 4
cases. Could someone please help me? Which, if any, of the estimates for
the intercept is the best one? And why are they different?

 

I think the error occurs because the MLE for the between group variance
is zero and somehow the estimate becomes negative. I appreciate any
help. 

 

Thanks,

 

Emine

 

 

install.packages("nlme")

install.packages("Matrix")

install.packages("lme4")

install.packages("car")

library(nlme)

library(Matrix)

library(lme4)

library(car)

library(arm)

 

 

y <- c(14, 9, 19, 12, 10, 12, 8, 11, 15, 4, 14, 13, 8, 3)

n <- c( 20, 20, 20, 20, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18)

center <- seq(1:14)

 

example1 <- lmer(cbind( y, n - y) ~ 1   + ( 1 | center) ,

family = binomial, niter = 50, method = "Laplace", control = list
(usePQL = FALSE))

 

fixef(example1)

se.fixef(example1)

 

example2 <- lmer(cbind( y, n - y) ~ 1   + ( 1 | center) ,

family = binomial, niter = 50, method = "Laplace", control = list
(usePQL = TRUE))

 

fixef(example2)

se.fixef(example2)

 

example3 <- lmer(cbind( y, n - y) ~ 1   + ( 1 | center) ,

family = binomial, niter = 50, method = "PQL",control = list (usePQL
TRUE))

 

fixef(example3)

se.fixef(example3)

 

example4 <- lmer(cbind( y, n - y) ~ 1   + ( 1 | center) ,

family = binomial, niter = 50, method = "PQL",control = list (usePQL
FALSE))

 

fixef(example4)

se.fixef(example4)

 

 

 

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~    

    Emine Ozgur BAYMAN
    PhD Student
    Department of Biostatistics, 
    College of Public Health,
    University of Iowa
    
    Office: 2400D UCC, Iowa City / IA
    Office Phone: (319) 384 27 52
    e-mail: emine-unlu@uiowa.edu

 


	[[alternative HTML version deleted]]