Skip to content
Prev 16109 / 20628 Next

Help - I have an underdispersed glmm :(

Hello Ben and everybody,

Thanks for your answer. I will try to answer to your questions in order.

Maybe I expressed myself wrong. I have no good reason to expect q1 to
follow poisson. All I understand - in my comic book type understanding-
is that shannon's H is a measure of diversity that results from the
weighting of the number of species in a sample by the individual
abundance within each species. Therefore q1 is a proportion and in that
case is not formally a count, so poisson will not work.

As to what sort of variance structure to expect from this response
variable, what I can say is that when I ignore the random effects and
fit a glm with gaussian (link="log"), there is almost no
heteroscedasticity, which is confirmed by a Bartlett's test
(below). However, the Q-Q plot of this same model shows that the std.dev
residuals do not adjust well to normality. I then re-specify the same
glm with a gamma distribution and log link function and that seems to
resolve both the heteroscedasticity and the fit of the std.dev
residuals.

Bartlett test of homogeneity of variances
data: resid(glm) by meta.ec.n$Treatment
Bartlett's K-squared = 0.94193, df = 3, p-value = 0.8153

Finally, I am not aware of any theoretical framework for the
distribution of this variable. And the reason for not simply
transforming my response variable to linearize the relationship and then
fit my data with lmer is that I was under the impression that GLMM are a
more efficient way to deal with these issues. The other reason is that I
am worried that the transformation over corrects the residual fit to
the theoretical distribution.

In addition to the statements of the glm, I ran a boxcox command (from
package:MASS) on the glm. It seemed to confirm that log transformation
of the response variable would be the best shot. I then specify a linear
mixed effect model, with the log transformed variable. However the
diagnostic plots, showed some heteroscedasticity and not that optimal fit to normality.

lme.1 <- lmer(log.q1~Treatment +(1|Elevation) + (1|Elevation:Plot), data=dat)

Do you think it would be better then to go for the linear model transforming the data, instead of the glmm?

Best regards,

Juan

/
/