Skip to content

Additive versus multiplicative overdispersion modeling

8 messages · Ned Dochtermann, Shinichi Nakagawa, Jarrod Hadfield +1 more

#
First posting to the list, prior to sending this out I've tried
searching the mixed model list, other lists and anything google could
pick up.

I am currently trying to calculate repeatability estimates
(intra-class correlation coefficients) following Nakagawa & Schielzeth
(2010, Biol.Rev. Repeatability for Gaussian and non-Gaussian data: a
practical guide for biologists. online early). The details of my
models shouldn't be important except that I originally fit the models
using binomial error structures and a logit link. Nakagawa and
Schielzeth (henceforth N&S) specify that repeatability estimates
differ based on whether additive or multiplicative overdispersion
modelling is conducted.

N&S define multiplicative as when the dispersion parameter is
estimated but that residual variance is fixed to one. Additive is
defined as having the residual variance estimated and the dispersion
parameter fixed to one. These definitions are based on Browne et al.
(2005, J. Roy. Stat. Soc A, 168:599-613).

Based on my reading of the family objects description it seems that
using the quasibinomial family would correspond to the multiplicative
overdispersion modelling and the binomial family would correspond to
additive overdispersion modelling.

Is this conclusion about multiplicative vs. additive correct or am I
missing something? I do realize that when the dispersion parameter is
estimated as being close to one under a quasibinomial model then the
results should be close to what you'd get with a binomial approach
which makes me think I am missing something (since the multiplicative
model would have the residual variance fixed).

Thank you for any help you can provide.
Ned Dochtermann



--
Ned Dochtermann
Department of Biology
University of Nevada, Reno
--
#
On Thu, 19 Aug 2010, Ned Dochtermann wrote:

            
[SNIP]
Yes.  Browne et al say they are using the "additive" approach because it 
has a proper likelihood.

If you are interested in repeatability of binary measures, there are lots 
of perfectly good "direct" measures.  The thing about the GLMM variance 
components is that they are up in the latent variable part of the model. 
If you are using a probit-normal, you are getting (essentially) 
tetrachoric correlations, that is, estimating the correlation between the 
"true" continuous measures that are being arbitrarily dichotomized to give 
you your binary outcome.  For biometrical geneticists, this is a regarded 
as a good thing (Yule might disagree ;)), but might not be as useful for, 
say, assessing different clinical tests.  It really does depend on your
actual problem.

Cheers, David Duffy.
#
Thanks a lot, if that is indeed the case it makes calculating
repeatabilities per N&S quite straightforward for the multiplicative
models (quasibinomial & quasipoisson) since the relevant term to
include in the denominator would just be (summary(model)@sigma)^2
(multiplied by (pi^2)/3 ). Of course I still can't figure out how to
get the needed information from the additive models, i.e. the residual
of the distribution specific variance.


Ned
On Thu, Aug 19, 2010 at 10:00 PM, David Duffy <davidD at qimr.edu.au> wrote:
#
Hi Ned,

You can get the additive residual term of N&S by fitting an  
observation-level random effect (i.e. one effect for each datum). You  
will need the latest version of lme4 for this (not available for Mac).  
If the data are binary you can't estimate the residual, so it is usual  
just to set it to zero.

Cheers,

Jarrod


Quoting Ned Dochtermann <ned.dochtermann at gmail.com>:

  
    
#
Hi, Jarrod

I think you can probably install lme4 on Mac - see the blog and its correspondences (I have not tried myself).

http://www.stat.columbia.edu/~cook/movabletype/archives/2010/08/multilevel_mode_11.html

Thanks for the tip

Best wishes

Shinichi

Shinichi Nakagawa, PhD
(Lecturer of Behavioural Ecology)
Department of Zoology
University of Otago
340 Great King Street
P. O. Box 56
Dunedin, New Zealand
Tel:  +64-3-479-5046
Fax: +64-3-479-7584
http://www.otago.ac.nz/zoology/staff/academic/nakagawa.html
#
Hi Shinichi,

Just tried - this works for me.

Cheers,

Jarrod
Quoting Shinichi Nakagawa <shinichi.nakagawa at otago.ac.nz>:

  
    
1 day later
#
On Fri, 20 Aug 2010, Ned Dochtermann wrote:

            
Method "C" in the Browne paper uses: r = V/(V+pi^2/3) for the logistic 
link, and r=V/(V+1) for the probit link (the latter is the tetrachoric r).

Cheers, David Duffy.
2 days later
#
David,

To somewhat wrap things up; as I guess would be expected, I get the same
repeatability estimate from a quasibinomial model using V/(V+sigma^2*pi^2/3)
as with V/(V+pi^2/3) from a binomial model.

Thanks again for your and everyone else's help!
Ned

--
Ned Dochtermann
Department of Biology
University of Nevada, Reno

ned.dochtermann at gmail.com
http://wolfweb.unr.edu/homepage/mpeacock/Dochter/
--



-----Original Message-----
From: David Duffy [mailto:davidD at qimr.edu.au] 
Sent: Sunday, August 22, 2010 6:39 PM
To: Ned Dochtermann
Cc: r-sig-mixed-models at r-project.org
Subject: Re: [R-sig-ME] Additive versus multiplicative overdispersion
modeling
On Fri, 20 Aug 2010, Ned Dochtermann wrote:

            
Method "C" in the Browne paper uses: r = V/(V+pi^2/3) for the logistic 
link, and r=V/(V+1) for the probit link (the latter is the tetrachoric r).

Cheers, David Duffy.