[R-sig-ME] Modelling random effects with SITE, YEAR and SPECIES

[CL Pressland]

Theirry,

thank you for you informative reply. I have had a go at your suggestions 
but have been stumped:

--On 07 May 2009 10:16 +0200 "ONKELINX, Thierry" <Thierry.ONKELINX at inbo.be> 
wrote:

> Dear Kate,
>
> Adding SPECIES as a random effect indicates that you want to take the
> effect of SPECIES into account but not need to know the effect of the
> individual SPECIES. If you do want to know that effect then you have to
> add species to fixed effects. Examining the effect of A, B and C on
> species (as a fixed effect) requires interactions between them. The
> model then looks like (A + B + C) * SPECIES + Year + (1|SITE) + (1|YEAR)
> This will only work if you have sufficiend data.

I tried this approach with data I have that is SPECIES recorded as SITES 
over YEARS but when I tried A*SPECIES as a fixed factor I received this 
error message:

"Error in mer_finalize(ans) : Downdated X'X is not positive definite, 88."

I've searched for what this error means but I cannot understand it.

This was written by Douglas Bates in response to [Re: [R] lme4, error in 
mer_finalize(ans)] posted 05 Dec 2008:
"That, admittedly obscure, error message relates to the fixed-effects 
specification rt ~ length + length:pos being rank deficient. If you look at 
the summary of the linear model fit you will see that there are 3 
coefficients that are not determined because of singularities. The lm 
function detects the singularities and fits a lower-rank model.  The lmer 
function is not as sophisticated. It just detects the singularities and 
quits."

I am unsure what this means or how it translates to my data. In my example, 
I have 78 "SPECIES" (factor, coded as numbers) and "A" is ordered data 0, 
1, 2. The y variable is number/m. You wrote that this would only work is 
you had sufficient data - each species is not recorded each time, so is 
this reduced data the cause i.e. not enough observations for n?

> Another option is to keep species as a random effect and add random
> slopes according to A, B and C. This will allow a different effect of A,
> B anc C for each species. The model would look like A + B + C + Year +
> (1|SITE) + (1|YEAR) + (A + B + C|SPECIES)

I have tried this way also but I am unsure of the output - it does not give 
species specific information and therefore I cannot work out which species 
is more affected by A, only if SPECIES as a whole are affected or not by 
each category of A. This is not useful to me as I would like to determine, 
given the random effects, if A 0, 1, or 2 affect which species in the data 
set.

Any thoughts?

Kate
>
> HTH,
>
> Thierry
>
> ------------------------------------------------------------------------
> ----
> ir. Thierry Onkelinx
> Instituut voor natuur- en bosonderzoek / Research Institute for Nature
> and Forest
> Cel biometrie, methodologie en kwaliteitszorg / Section biometrics,
> methodology and quality assurance
> Gaverstraat 4
> 9500 Geraardsbergen
> Belgium
> tel. + 32 54/436 185
> Thierry.Onkelinx at inbo.be
> www.inbo.be
>
> To call in the statistician after the experiment is done may be no more
> than asking him to perform a post-mortem examination: he may be able to
> say what the experiment died of.
> ~ Sir Ronald Aylmer Fisher
>
> The plural of anecdote is not data.
> ~ Roger Brinner
>
> The combination of some data and an aching desire for an answer does not
> ensure that a reasonable answer can be extracted from a given body of
> data.
> ~ John Tukey
>
> -----Oorspronkelijk bericht-----
> Van: r-sig-mixed-models-bounces at r-project.org
> [mailto:r-sig-mixed-models-bounces at r-project.org] Namens CL Pressland
> Verzonden: woensdag 6 mei 2009 20:18
> Aan: R Mixed Models
> Onderwerp: Re: [R-sig-ME] Modelling random effects with SITE, YEAR and
> SPECIES
>
> How can you work out how A, B or C affect SPECIES? By this I mean, could
>
> you find out how species n is affected by A, B and C in the correlation
> output? Or would you need to adjust the response to look at individual
> species separately?
>
> --On 29 April 2009 17:58 -0400 Ben Bolker <bolker at ufl.edu> wrote:
>
>> David R. wrote:
>>> Hello all,
>>>
>>> First, sorry for the english and the basic questions. I'm using mixed
>>> models (lme4 package) to analyse variability in 13 SPECIES of birds
>>> observed during 15 YEARS across 5 SITES. All the SPECIES were
>>> observed in all the sites in most years.
>>>
>>> My fixed effects are A, B, C and Year. I'm interested in the
>>> stochastic effect of A, B and C on the dependent variable, but also
>>> in a possible linear trend of the dependent variable over time.
>>>
>>> My random effects are SPECIES, YEAR and SITE, to control for the
>>> effects of nonindependence.
>>>
>>> I have a model with SITE, YEAR and SPECIES as crossed random effects
>>> like A + B + C + Year + (1|SITE) + (1|YEAR) + (1|SPECIES).
>>>
>>> My questions are:
>>>
>>> 1) Is this model correct? It is correct to model YEAR both as random
>>> effect and fixed effect? Is there the possibility that the variance
>>> accounted for by the random effect could robbing year as a fixed
>>> effect of explanatory power?
>>
>>   Seems OK and sensible to me.
>>   I would guess that the linear trend and the random variation are
>> sufficiently different patterns that they would not conflict too
> badly,
>> but you could try the different nested models and see what happens ...
>>
>>>
>>> 2) It is meaningful, instead,  to model YEAR as repeated measure, if
>>> the experimental unit were species within sites?
>>
>>   "Modeling YEAR as a random effect" and "Modeling YEAR as a repeated
>> measure" are, in my opinion, almost the same thing (but I'm ready to
> be
>> corrected, as always).  The only aspect of "repeated measures" that
>> would be different would be if you wanted to fit an autoregressive
> model
>> so that samples closer together in time were more correlated (which
> you
>> can't do with lmer at this
>> point).
>>
>>   Ben Bolker
>>
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>
>
> ----------------------
> Kate Pressland
> Office D95
> School of Biological Sciences
> University of Bristol
> Woodland Road
> Bristol, BS8 1UG
> Tel: 0117 9288918 (Internal 88918)
> Kate.Pressland at bristol.ac.uk
> www.bio.bris.ac.uk/people/staff.cfm?key=1137
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
> Dit bericht en eventuele bijlagen geven enkel de visie van de schrijver
> weer  en binden het INBO onder geen enkel beding, zolang dit bericht niet
> bevestigd is door een geldig ondertekend document. The views expressed in
> this message  and any annex are purely those of the writer and may not be
> regarded as stating  an official position of INBO, as long as the message
> is not confirmed by a duly  signed document.



----------------------
Kate Pressland
Office D95
School of Biological Sciences
University of Bristol
Woodland Road
Bristol, BS8 1UG
Tel: 0117 9288918 (Internal 88918)
Kate.Pressland at bristol.ac.uk
www.bio.bris.ac.uk/people/staff.cfm?key=1137