[R-sig-ME] R: Re: glmmadmb and time-effects

[Ben Bolker]

[cc'ing back to r-sig-mixed-models]

On 13-08-15 09:20 AM, Luca Corlatti wrote:
> Dear Ben, 
> sorry for the late reply. Thanks for your advice, that's exactly what I
> needed. The inclusion of the random effect (variable|year:month) might
> be a bit tricky as several variables (cortisol, testosterone and home
> range, for instance) may be collinear with time, and I guess handling so
> many random factors may be difficult?

  Maybe.  It might be worth a try though.  Depends on the size of your
data set.

> Apart from this, given the confounding effect of time, would it be
> correct to assume that the model selection with AIC may help anyway to
> disentangle to relative role of my variables to explain the parasite
> emission pattern? 

  Yes, although see the various caveats at http://glmm.wikidot.com/faq
(search for "AIC")

> Do you suggest in the discussion of my work I shall
> state some variables may be somewhat cofounded with time?

  Yes.

> Thanks in advance, 


> Luca
>  
> 
>>>> Ben Bolker <bbolker at gmail.com> 11/08/13 20.15 >>>
> Luca Corlatti <luca.corlatti at ...> writes:
> 
>>
>> Dear all,
> 
>> I am trying to analyse the relationship between parasite burden and
>> several internal and external variables, including testosterone,
>> cortisol, age, minimum temperature, home range. I have 2 years of
>> data, collected on a monthly basis. My data are not normally
>> distributed and overdispersed. I therefore fitted my global model
>> as:
> 
> mod <- glmmadmb(parasite~testosterone + cortisol + age + Tmin +
> hr + age:testosterone + age:cortisol + (1|year:month) + (1|id),
> family="nbinom", data=mydata, ZeroInflation=FALSE)
> 
>> Visual inspection of residuals suggest that the model fits the data
>> adequately. Starting from here, I fitted a set of simpler models
>> and ran a model selection and a model averaging of the competitive
>> models.
> 
>> The parasite emission shows marked monthly variation but, clearly,
>> all the independent variables as somewhat dependent on time as well,
>> and if I included month (time) as a fixed factor in the model, I am
>> afraid the effects of such variables would be diluted. I therefore
>> decided to include time as a random factor (1|year:month), but I am
>> not sure if this is a plausible choice. Kind regards, Luke
> 
> It seems plausible, although technically if your independent variables
> are collinear with time, the most conservative/honest thing to do is
> to admit that your variables of interest are somewhat confounded with
> time. In other words, including time as a random factor should
> dilute the effect slightly less than including it as a fixed factor,
> but it will still dilute it some, and you can't really get around that.
> 
> If some of your variables take on more than one value at each time
> step, you might consider including their interactions with time, e.g.
> (cortisol|year:month) to allow for variation in the _effect_ over time
> (the intercept-by-time model only allows for variation in the baseline
> parasite emission over time) -- see e.g. Schielzeth and Forstmeier
> Behavioural Ecology 2009 ...
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>