Hi everyone,
I have posted a similar question to this list, but I don't
get a reply. I really want to solve this problem, so I
post it again...
I am trying to use R to fit some mixed-effects models for
a nested data. The data is a simulated data with 111
subjects. Each subject has 6 waves' data. Below are the
first two subjects' data : > simu1[1:12,]
Grouped Data: gf ~ age | id
id occasion age gf tau b0
b1 b2
1 1 1 4.056714 35.08943 14.8 14.502078
5.209333 -5.199452
2 1 2 14.056714 87.18276 14.8 14.502078
5.209333 -5.199452
3 1 3 29.056714 91.19566 14.8 14.502078
5.209333 -5.199452
4 1 4 42.056714 91.32411 14.8 14.502078
5.209333 -5.199452
5 1 5 57.056714 91.47233 14.8 14.502078
5.209333 -5.199452
6 1 6 65.056714 91.55138 14.8 14.502078
5.209333 -5.199452
7 2 1 3.353627 16.94166 14.8 -0.331536
5.293478 -5.392813
8 2 2 13.353627 69.87644 14.8 -0.331536
5.293478 -5.392813
9 2 3 28.353627 76.18644 14.8 -0.331536
5.293478 -5.392813
10 2 4 41.353627 74.89508 14.8 -0.331536 5.293478
-5.392813
11 2 5 56.353627 73.40505 14.8 -0.331536 5.293478
-5.392813
12 2 6 64.353627 72.61037 14.8 -0.331536 5.293478
-5.392813
Firstly, I used lme to fit a mixed effects model with
quadratic curve.> simu1$age2<-simu1$age^2
> mm.lme.2<-lme(gf~age+age2,random=~age+age2|id,data=simu1)
It works well:> summary(mm.lme.2)
Linear mixed-effects model fit by REML
Data: simu1
AIC BIC logLik
5222.225 5267.193 -2601.113
Random effects:
Formula: ~age + age2 | id
Structure: General positive-definite, Log-Cholesky
parametrization
StdDev Corr
(Intercept) 2.729384e+00 (Intr) age
age 1.068434e-01 0.648
age2 8.285027e-34 0.000 0.000
Residual 1.082344e+01
Fixed effects: gf ~ age + age2
Value Std.Error DF t-value p-value
(Intercept) 23.902926 1.1606163 553 20.59503 <.0001
age 3.046429 0.0815675 553 37.34854 <.0001
age2 -0.034015 0.0011395 553 -29.85176 <.0001
Correlation:
(Intr) age
age -0.790
age2 0.697 -0.964
Standardized Within-Group Residuals:
Min Q1 Med Q3 Max
-2.5233682 -0.6801665 -0.1735055 0.5243569 2.7772921
Number of Observations: 666
Number of Groups: 111
However, when I tried to fit the following model and the
starting value is very close to the true value:> simu.nlme<-nlme(gf~b00 + b10 * age + b20 *
max(0,(age-tau0)),data=simu1,fixed=b00+b10+b20+tau0~1,random=b00+b10+b20+tau0~1,group=~id,
start=c(b00=4.08,b10=5.32,b20=-5.29,tau0=14.8),method="REML")
It shows following error:
**************************************
Error in MEEM(object, conLin, control$niterEM) :
Singularity in backsolve at level 0, block 1
**************************************
I also fitted the same model by using SAS and winBUGS.
They all worked well...
Do you have any suggestions on this problem?
Any reply will be appreciated!
Lijuan
