Zero variance and Std. Dev. using lmer?
On Fri, 16 Jan 2009, Jude Phillips wrote:
Hi, I have a similar problem to Luciano. I am running a mixed effects model with a continuous dependent variable, a categorical fixed effect and a categorical random effect. The dependent variable has a distribution that is skewed to the left, which can be normalized by a log transformation. There is a huge difference in the results I get, depending on whether I use the transformation, and I am not sure why. spimelog<-lmer(log(X13c +28) ~ crop + (1|field.single), spi)
From the previous posts, I think I understand that the first model has
0 std.dev and var for the random effects because the log-likelihood is not being evaluated correctly, but why is the result so different when the dependent variable is transformed. (note that the dependent variable happens to take negative values, the lowest of which is -27.5, which is why I add 28 before the log transformation).
Because these types of analysis are sensitive to the distribution of y. Have you looked at the distribution of the field means under the two transformations? Not knowing (and not wanting to know ;)) about your data, is log(x + c) with c=28 a bit extreme? Have you looked at Box-Cox type approaches? David Duffy | David Duffy (MBBS PhD) ,-_|\ | email: davidD at qimr.edu.au ph: INT+61+7+3362-0217 fax: -0101 / * | Epidemiology Unit, Queensland Institute of Medical Research \_,-._/ | 300 Herston Rd, Brisbane, Queensland 4029, Australia GPG 4D0B994A v