Some questions to GLMM
Jan Hattendorf wrote:
Hello all R-user I am relative new to the R-environment and also to GLMM, so please don't be irritated if some questions don't make sense. I am using R 2.0.0 on Windows 2000. I investigated the occurrence of insects (count) in different parts of different plants (plantid) and recorded as well some characteristics of the plant parts (e.g. thickness). It is an unbalanced design with 21 plants with approx 25 parts each. Preference of the insects for a certain characteristic is usually unimodal. As far as I understood, I have to use a model with random intercepts and slopes, because the observations within each plant are not independent. So far so good ========(lme4)========= glmm1<-GLMM(count~thick+I(thick^2),random=~thick+I(thick^2) |plantid,poisson,data=Dataset,control=list(PQLmaxIt=10000))
summary(glmm1)
Generalized Linear Mixed Model
Family: poisson family with log link
Fixed: lixt ~ thick + I(thick^2)
Data: Dataset
AIC BIC logLik
-125.2406 -83.7346 72.62031
Random effects:
Groups Name Variance Std.Dev. Corr
plantid (Intercept) 0.0173455 0.131702
thick 0.0389772 0.197426 -1.000
I(thick^2) 0.0013327 0.036507 1.000 -1.000
# of obs: 469, groups: plantid, 21
Estimated scale (compare to 1) 1.402567
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -4.045569 0.346950 -11.660 < 2.2e-16 ***
thick 2.378207 0.195425 12.169 < 2.2e-16 ***
I(thick^2) -0.280898 0.025458 -11.034 < 2.2e-16 ***
---
Signif. codes: 0 `***' 0.001 `**' 0.01 `*' 0.05 `.' 0.1 ` ' 1
Correlation of Fixed Effects:
(Intr) thick
thick -0.968
I(thick^2) 0.900 -0.977
============================
Question 1: Is the formula suitable for my design?
My guess is that your model is overparameterized. Notice that the estimated correlations of the random effects are all near the extremes of -1 or +1. I would start with a model that had fewer random effects terms.
Question 2: What can be the reason for the positive logLik-value?
It is a common misconception that a log-likelihood must be negative. In fact, a log-likelihood can be positive when it is based on a probability density. Probabilities cannot exceed one but probability densities can. In this case the log-likelihood is a combination of the probability density for the random effects and the conditional probability of the observations.
Question 3: It is correct, that I do not have to take care about over- dispersion.
There is an indication of overdispersion but I would not worry about that until I could get a handle on the random effects terms.
Question 4: When I use "poly(thick,2)" instead of "thick+I(thick^2)" I get completly different estimate-values (the latter one are the correct one). I thought it should be the same.
They fit the same model but with a different set of coefficients hence the estimated values are different.
============================
glmm2<-GLMM(count~poly(thick,2),random=~poly(thick,2)
|plantid,poisson,data=Dataset,control=list(PQLmaxIt=10000))
summary(glmm2)
[...]
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.69293 0.10711 -6.4694 9.837e-11 ***
poly(thick, 2)1 -47.23211 5.97336 -7.9071 2.634e-15 ***
poly(thick, 2)2 -51.21421 4.64972 -11.0145 < 2.2e-16 ***
============================
Question 5: If I use the same formula in glmmPQL, I get more or less similar
results, but different values for AIC BIC and logLik.
I read in this thread:
http://maths.newcastle.edu.au/~rking/R/help/03b/6849.html
That it should be the same value due to the same algorithm
The GLMM function evaluates the likelihood at the PQL estimates using the Laplacian approximation so the result will be different from that returned by glmmPQL.
Maybe as additional comment, I specified a "NULL-model"
dummy<-rep(1,469) glmm3<-GLMM(count~1,random=~1|dummy,poisson)
resp
Dataset$dummy<-1 glmm3<-glmmPQL(count~1,random=~1|dummy,poisson)
and GLMM gave the correct value (logLik = (Null deviance/2) from glm),whereas glmmPQL calculated a logLik almost twice as high. =========(MASS)=====
glmm1<-glmmPQL(lixt~thick+I(thick^2),random=~thick+I(thick^2)
|plantid,poisson,Dataset) iteration 1 [...] iteration 10
summary(glmm1)
Linear mixed-effects model fit by maximum likelihood
Data: Dataset
AIC BIC logLik
1988.500 2030.006 -984.2502
Random effects:
Formula: ~thick + I(thick^2) | plantid
Structure: General positive-definite, Log-Cholesky parametrization
StdDev Corr
(Intercept) 0.27914612 (Intr) thick
thick 0.03089333 -0.251
I(thick^2) 0.01867846 -0.878 -0.067
Residual 1.37952490
Variance function:
Structure: fixed weights
Formula: ~invwt
Fixed effects: lixt ~ thick + I(thick^2)
Value Std.Error DF t-value p-value
(Intercept) -4.038800 0.4834792 446 -8.353617 0
thick 2.371842 0.2642474 446 8.975837 0
I(thick^2) -0.279189 0.0337451 446 -8.273479 0
Correlation:
(Intr) thick
thick -0.970
I(thick^2) 0.896 -0.973
Standardized Within-Group Residuals:
Min Q1 Med Q3 Max
-1.2093444 -0.5972699 -0.3725736 0.2819543 6.1820947
Number of Observations: 469
Number of Groups: 21
=============================
Question 6: When I tried method="Laplace" an error occurred:
g1<-GLMM(count~thick+I(thick^2),random=~thick+I(thick^2)
|plantid,poisson,data=Dataset,method="Laplace") Using optimizer nlm Error: rank deficiency of ZtZ+W detected at column 11
Is that indicating multicollinearity?
No, it is an indication that the variance-covariance of the random effects is not positive definite.
Is there is a possibility to avoid this?
You will need to reduce the number of random effects terms.
I got also from time to time the message Omega[1] is not positive definite Can I fix this somehow?
Same as above.
Question 7: Is there a possibility to calculate something like a pseudo-r- square (e.g. 1-(logLik(Nullmodel)/loglike model)?) Question 8: When I specify the whole model as fix as e.g: glm1<-glm(count=(thick+I(thick^2))+(thick+I(thick^2))%in% plantid,poisson,data=Dataset) Is the model than wrong or just less powerful? I guess the latter is the case, but I would like to be sure. If in this version is just the power decreased, it would be very helpful for me to use for the more complex models this approach, because errors and warnings are becoming more frequent with more factors and covariates. I know that there are a lot of questions (I am sorry for that), and I don't expect that all will be commented. Most interesting for me are the questions 1,6 and 8. Thank you very much in advance. Jan ======================================== Jan Hattendorf University of Berne Zoological Institute Baltzerstrasse 6 CH-3012 Berne +41-31-631 4523 jan.hattendorf at zos.unibe.ch http://www.cx.unibe.ch/zos/index.html
______________________________________________ R-help at stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html