[R-sig-ME] MCMCglmm ZIP longitudinal

[Thomas Houslay]

Hi David, thanks for the response - I very much appreciate any help that's on offer, no matter how general! At the very least, it's nice to know that my question on mixing is 'very hard' and I haven't been staring at plots etc entirely without cause...

As far as the interaction terms go, I think I can get rid of those involving lifespan, but the diet treatments and age-related terms (plus interactions) are necessary, I think - every individual has been assigned to larval and adult treatment, and I want to see whether the pattern of age-related calling is affected by treatment group...

Anyway, I shall take a look at what you advise at the end of your email, although unfortunately I evidently have to read up on those type of models before it's clear to me what to do!

Thanks again,

Tom


-----Original Message-----
From: David Duffy [mailto:David.Duffy at qimr.edu.au] 
Sent: 29 July 2013 01:46
To: Thomas Houslay
Cc: r-sig-mixed-models at r-project.org
Subject: Re: [R-sig-ME] MCMCglmm ZIP longitudinal

On Sat, 27 Jul 2013, Thomas Houslay wrote:

> I have been struggling with some analysis for quite a long time now, 
> and hoping someone can help give me a little advice as to whether my 
> current model (using MCMCglmm) seems reasonable, in addition to 
> answering a question about autocorrelation in the zero-inflation term.
>
> The responses are counts of the time that an individual cricket spent 
> calling on the day of measurement (each individual is measured at 
> weekly intervals from 1 week after reaching adulthood until death), 
> and appear to be highly overdispersed compared to standard Poisson, 
> and also zero-inflated - I have therefore been using the ZIP model in MCMCglmm.
[...]
> Each individual was raised on one of two diet treatments during the 
> nymph stage (ie, before measurements started), and then assigned to 
> one of two diet treatments again at adulthood. I am interested in the 
> effect of diet and age (and their interactions) on the pattern of 
> calling.

I have spent some time resisting the urge to respond to this :), since I can make only really general remarks. The question about mixing is really hard - I see these models and methods as still experimental.  The proposals for the MCMC may be exploring the whole space really slowly because the parameterization is not optimal for that particular problem, or it can just be stuck. There are no foolproof diagnostics, so experts still spend a lot of time staring at plots, and simulating data from the proposed model and seeing if they then can recapture the right answer.

And I have great difficulty understanding what you will do with all these interaction terms.  If you can't get rid of them my first thought would be to maybe try a "straightforward" non-linear or semiparametric (GAMMs etc) model for individual trajectory, and look for treatment effects on velocity or higher order terms.

Just 1.5c, David.

| 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


-- 
The University of Stirling is ranked in the top 50 in the world in The Times Higher Education 100 Under 50 table, which ranks the world's best 100 universities under 50 years old.
The University of Stirling is a charity registered in Scotland, 
 number SC 011159.