Hi all! I hope somebody can help me solve some doubts which must be very basic, but I haven't been able to solve by myself. The first one, is how to assess for overdispersion in GlmmPQL when fitting binomial or poisson errors. The second one is whether GlmmPQL can compare models with different fixed effects. The third doubt, regards the way I should arrange my data in a GlmmPQL with binomial errors. In glm, I am supposed to create cbind vector joining the "number of successes" and the "total-the number of successes". Should I proceed the same way for GlmmPQL of can I use a single column which, intead of containing the numbers, simply contains 0 or 1?. The reason for this question, is that I am trying to fit a variance components analysis with a single random effect and no fixed effects. The only way I know to test for the significance of the single level of random effects is by comparing the model with a glm without fixed effects and do a ChiSquare test. So, should the data of both models be arranged the same way? or is it possible to compare the model with random effects and response "0,1" whith that of a glm without fixed effects where the response is arranged as cbind(successes,total-successes)? My concern of using cbind in GlmmPQL, is that lose replication of the variable of interest, and I have realized that when fitting the random variable as a fixed effect in a simple glm, I end up with 0 residual degrees of freedom. Thanks in advance for your help, Olga
glmmPQL is part of MASS, and that is support software for a book. Do look in the book and its references .... On Sat, 6 Mar 2004, O Tosas Auguet wrote:> I hope somebody can help me solve some doubts which must be very basic, > but I haven't been able to solve by myself. > > The first one, is how to assess for overdispersion in GlmmPQL when fitting > binomial or poisson errors. The second one is whether GlmmPQL can compare > models with different fixed effects. > > The third doubt, regards the way I should arrange my data in a GlmmPQL with > binomial errors. In glm, I am supposed to create cbind vector joining > the "number of successes" and the "total-the number of successes". Should I > proceed the same way for GlmmPQL of can I use a single column which, intead of > containing the numbers, simply contains 0 or 1?. The reason for this question, > is that I am trying to fit a variance components analysis with a single random > effect and no fixed effects. The only way I know to test for the significance > of the single level of random effects is by comparing the model with a glm > without fixed effects and do a ChiSquare test. So, should the data of both > models be arranged the same way? or is it possible to compare the model with > random effects and response "0,1" whith that of a glm without fixed effects > where the response is arranged as cbind(successes,total-successes)? My concern > of using cbind in GlmmPQL, is that lose replication of the variable of > interest, and I have realized that when fitting the random variable as a fixed > effect in a simple glm, I end up with 0 residual degrees of freedom.-- Brian D. Ripley, ripley at stats.ox.ac.uk Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UK Fax: +44 1865 272595
At 13:52 2004-03-06, you wrote:>The reason for this question, is that I am trying to fit a variance >components analysis with a single random effect and no fixed effects.The >only way I know to test for the significance of the single level of random >effects is by comparing the model with a glm without fixed effects and do >a ChiSquare test.This may or may not be known to you: If one model is a special case of the other, i.e. its variance component equals zero, then the less complex model falls on the boundary of the parameter space relative to the more complex model. Hence the LR test needs to be adjusted, and in this simple case the adjusted p-value may be taken as half the p-value from the usual one degree of freedom chi-squared test, i.e. using a 50-50 mixture of chi-squared distributions with zero and one df as the reference. In other cases, e.g. when simultaneously testing if two or more variance components equals zero, the asymptotic distribution isn't generally known. You may use a score test (Lin, Biometrika, 1997), but I don't think it's implemented in R or any of the add-on packages. Pinheiro & Bates "Mixed-effects models in S and S-PLUS" p. 82-87 provide a good discussion in the context of linear mixed-effects models. //Henric