[R-sig-ME] heteroscedastic model in lme4

Rense Nieuwenhuis contact at rensenieuwenhuis.nl
Thu Jan 15 14:00:39 CET 2009

Dear all,

I'm currently working on a similar problem in which I expect the  
within-group variance to change over years. Of course, in an ideal  
world correlation structures would be readily available in lme4, but  
as Thierry already states, this is not the case.

I have been considering a different approach to solve the problem,  
based on statements on heteroscedasticity made by Snijders & Bosker  
(1999, page 119). This only works for heteroscedasticity at the second  
(or higher) level, and if this heteroscedasticity can be related to an  
observed variable.

He argues that the random part of the second level can be defined as:
U0j + U1j*zj
"Thus, strange as it may sound, the level-two variable Z formally gets  
a random slope at level two".

In my understanding, this would lead to something like:

glmer(Count ~ A + B + (B|Group), family=poisson)

assuming that B measures a  'group'-level characteristic.

I'm still trying to get a firm grasp on the matter, but possibly this  
solution can help out in certain situations,

kind regards,

Rense Nieuwenhuis

On 15 jan 2009, at 11:13, ONKELINX, Thierry wrote:

> glmer(Count ~ A + B + (1|Group), family = poisson)

	Author = {Snijders, Tom A.B. and Bosker, Roel J.},
	Date-Added = {2009-01-15 13:48:53 +0100},
	Date-Modified = {2009-01-15 13:49:35 +0100},
	Publisher = {Sage},
	Title = {Multilevel Analysis, an introduction to basic and advanced  
multilevel modelling},
	Year = {1999}}

More information about the R-sig-mixed-models mailing list