Dear list members, I constructed this model: bao1<-lme(sla~mg, random=~pop/nr.tree, weights=varPower(form=~sla|pop) variables: - sla = continuus - mg = factor - pop = factor - nr.tree = factor So, the variance of sla increases with sla, dependent of the pop. However, I fitted another (homoscedastic) model, where I transformed sla to log(sla): bao2<-lme(log(sla)~mg, random=~pop/nr.tree) This second model predicts in a better way the observations than the first model (observed with a sla against fitted(.) plot). Which model is the most convenient? Are there advantages/disadvantages in the use of these models? Thanks in advance, Sebastiaan De Smedt Department of Bioscience Engineering University of Antwerp Belgium Tel.: +32 (0)3 265 35 17 Fax.: +32 (0)3 265 32 25 [[alternative HTML version deleted]]