Quantcast

mixed effects in ordination (?)

classic Classic list List threaded Threaded
3 messages Options
Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

mixed effects in ordination (?)

Martin Weiser
Dear friends,

Could you please help me with analysis? I am afraid that it involves
crossed random effects in the mixed-effect constrained ordination
setting, so to say.

Goal:
Show an effect of the species trait (single one) and treatment (four
levels, quantitative scale) on parameters. Trait x treatment
interaction is possible. If possible, show changes through time.

Data:
Individuals of 7 species, subjected to 4 treatment levels (fully
factorial) - from 6 to 12 individuals in each combination. Each
individual scored 4 times (same times: 1st wk, 2nd wk, 3rd wk, 4th wk).
Several (10) parameters scored every time on each individual.

What I did:
In order to avoid multiple testing (the parameters are likely to be
correlated to each other), I decided to use multivariate analysis
(RDA). I am by far much more accustomed to vegan than ade4, so excuse
me if I use some "veganisms". Predictors: time, trait, treatment
(possibly with interactions), conditioned on individual identity to
avoid treating records from the same plant as independent. Variance
partitioning.

Here comes the problem: how to set permutations in order to be able to
report p_vals (some people just are not happy without them)? Since
individuals of the same species share the same trait value, maybe the
right way is to: shift observations within individual (if time is among
predictors for the particular model) and permute trait value among
species. Is it so? Is this treatment of the pseudoreplication at the
species level (i.e. only in the significance testing, not in the
ordination per se) ok?

I also tried to use different approach: I averaged all params
individual-wise (getting rid of temporal pseudoreplication, but also
time effects), and subsequently I averaged the result within treatment
x species levels. I assume that I can go for simple free permutations
this way? Pity is that this way, I cannot see development in time.

And another way: I averaged params for species x treatment x time
groups, ignoring interdependence of records from the same individual,
hoping that the effect of an individual "dissolves" in the average. Is
that meaningful? If yes, what is the appropriate permutation structure
in this case?

But maybe I miss something and there are better ways how to deal with
this problem?

Any suggestions (ok: not any, just those made in an attempt to help :-)
) are appreciated.

With the best regards,
Martin Weiser

_______________________________________________
R-sig-ecology mailing list
[hidden email]
https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: mixed effects in ordination (?)

Bob O'Hara
This doesn't sound like an ordination problem: it's more similar to a
standard mixed model regression. There are several ways to approach
this, for example Ben Bolker has some (old) notes here:
<https://rpubs.com/bbolker/3336>

I'd suggest you start by running the analyses on each trait
individually, as that way you can make sure you have the model sorted
correctly. The step up to the multivariate model just needs some work
understanding what the relevant R package wants.

HTH

Bob

On 12/01/17 14:12, Martin Weiser wrote:

> Dear friends,
>
> Could you please help me with analysis? I am afraid that it involves
> crossed random effects in the mixed-effect constrained ordination
> setting, so to say.
>
> Goal:
> Show an effect of the species trait (single one) and treatment (four
> levels, quantitative scale) on parameters. Trait x treatment
> interaction is possible. If possible, show changes through time.
>
> Data:
> Individuals of 7 species, subjected to 4 treatment levels (fully
> factorial) - from 6 to 12 individuals in each combination. Each
> individual scored 4 times (same times: 1st wk, 2nd wk, 3rd wk, 4th wk).
> Several (10) parameters scored every time on each individual.
>
> What I did:
> In order to avoid multiple testing (the parameters are likely to be
> correlated to each other), I decided to use multivariate analysis
> (RDA). I am by far much more accustomed to vegan than ade4, so excuse
> me if I use some "veganisms". Predictors: time, trait, treatment
> (possibly with interactions), conditioned on individual identity to
> avoid treating records from the same plant as independent. Variance
> partitioning.
>
> Here comes the problem: how to set permutations in order to be able to
> report p_vals (some people just are not happy without them)? Since
> individuals of the same species share the same trait value, maybe the
> right way is to: shift observations within individual (if time is among
> predictors for the particular model) and permute trait value among
> species. Is it so? Is this treatment of the pseudoreplication at the
> species level (i.e. only in the significance testing, not in the
> ordination per se) ok?
>
> I also tried to use different approach: I averaged all params
> individual-wise (getting rid of temporal pseudoreplication, but also
> time effects), and subsequently I averaged the result within treatment
> x species levels. I assume that I can go for simple free permutations
> this way? Pity is that this way, I cannot see development in time.
>
> And another way: I averaged params for species x treatment x time
> groups, ignoring interdependence of records from the same individual,
> hoping that the effect of an individual "dissolves" in the average. Is
> that meaningful? If yes, what is the appropriate permutation structure
> in this case?
>
> But maybe I miss something and there are better ways how to deal with
> this problem?
>
> Any suggestions (ok: not any, just those made in an attempt to help :-)
> ) are appreciated.
>
> With the best regards,
> Martin Weiser
>
> _______________________________________________
> R-sig-ecology mailing list
> [hidden email]
> https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
>


--
Bob O'Hara

NEW EMAIL: [hidden email]
NOTE NEW ADDRESS!!!
Institutt for matematiske fag
NTNU
7491 Trondheim
Norway

Mobile: +49 1515 888 5440
Journal of Negative Results - EEB: www.jnr-eeb.org

_______________________________________________
R-sig-ecology mailing list
[hidden email]
https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: mixed effects in ordination (?)

Mgr. Martin Weiser
Dear Bob and others,

Thank you very much for your answer. I realized I forgot to mention
the nature of the params: these are mostly root system architectural
traits, and, as you can imagine, they are just a subset of many
possible ones. The goal is merely to show how groups of traits are
related to the predictors.

In the terms of Ben Bolker's note you kindly pointed to me, imagine
that apart of cars and holiday, there will be also swimming pools,
paintings, houses and opera visits, restaurant dinners, charity
donations.  The question is then to show whether increase in
individual wealth is more pronounced in obtaining possessions or
immaterial "feelgoods". Maybe cars tightly correlate with swimming
pools and paintings and do not correlate so tightly with an individual
income,  feelgoods are correlated to each other and strongly predicted
by (correlated to) the income, while houses do not correlate with
anything (even though one would guess that they should behave
similarly to e.g. cars). If someone forgets to take notes about e.g.
paintings, nothing serious happens. P_val is there just to say that
the structure of the spendings is probably not random regarding to the
income.

I hope this made the problem clearer to anyone wishing to join the thread.

With the best regards,
Martin W.


2017-01-12 14:29 GMT+01:00 Bob O'Hara <[hidden email]>:

> This doesn't sound like an ordination problem: it's more similar to a
> standard mixed model regression. There are several ways to approach this,
> for example Ben Bolker has some (old) notes here:
> <https://rpubs.com/bbolker/3336>
>
> I'd suggest you start by running the analyses on each trait individually, as
> that way you can make sure you have the model sorted correctly. The step up
> to the multivariate model just needs some work understanding what the
> relevant R package wants.
>
> HTH
>
> Bob
>
> On 12/01/17 14:12, Martin Weiser wrote:
>>
>> Dear friends,
>>
>> Could you please help me with analysis? I am afraid that it involves
>> crossed random effects in the mixed-effect constrained ordination
>> setting, so to say.
>>
>> Goal:
>> Show an effect of the species trait (single one) and treatment (four
>> levels, quantitative scale) on parameters. Trait x treatment
>> interaction is possible. If possible, show changes through time.
>>
>> Data:
>> Individuals of 7 species, subjected to 4 treatment levels (fully
>> factorial) - from 6 to 12 individuals in each combination. Each
>> individual scored 4 times (same times: 1st wk, 2nd wk, 3rd wk, 4th wk).
>> Several (10) parameters scored every time on each individual.
>>
>> What I did:
>> In order to avoid multiple testing (the parameters are likely to be
>> correlated to each other), I decided to use multivariate analysis
>> (RDA). I am by far much more accustomed to vegan than ade4, so excuse
>> me if I use some "veganisms". Predictors: time, trait, treatment
>> (possibly with interactions), conditioned on individual identity to
>> avoid treating records from the same plant as independent. Variance
>> partitioning.
>>
>> Here comes the problem: how to set permutations in order to be able to
>> report p_vals (some people just are not happy without them)? Since
>> individuals of the same species share the same trait value, maybe the
>> right way is to: shift observations within individual (if time is among
>> predictors for the particular model) and permute trait value among
>> species. Is it so? Is this treatment of the pseudoreplication at the
>> species level (i.e. only in the significance testing, not in the
>> ordination per se) ok?
>>
>> I also tried to use different approach: I averaged all params
>> individual-wise (getting rid of temporal pseudoreplication, but also
>> time effects), and subsequently I averaged the result within treatment
>> x species levels. I assume that I can go for simple free permutations
>> this way? Pity is that this way, I cannot see development in time.
>>
>> And another way: I averaged params for species x treatment x time
>> groups, ignoring interdependence of records from the same individual,
>> hoping that the effect of an individual "dissolves" in the average. Is
>> that meaningful? If yes, what is the appropriate permutation structure
>> in this case?
>>
>> But maybe I miss something and there are better ways how to deal with
>> this problem?
>>
>> Any suggestions (ok: not any, just those made in an attempt to help :-)
>> ) are appreciated.
>>
>> With the best regards,
>> Martin Weiser
>>
>> _______________________________________________
>> R-sig-ecology mailing list
>> [hidden email]
>> https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
>>
>
>
> --
> Bob O'Hara
>
> NEW EMAIL: [hidden email]
> NOTE NEW ADDRESS!!!
> Institutt for matematiske fag
> NTNU
> 7491 Trondheim
> Norway
>
> Mobile: +49 1515 888 5440
> Journal of Negative Results - EEB: www.jnr-eeb.org



--
Mgr. Martin Weiser, Ph.D.
Tel.: +420 221 95 1654
E-mail: [hidden email]

Katedra botaniky; Přírodovědecká fakulta
Univerzita Karlova v Praze
Přírodovědecká fakulta
Benátská 2, 128 43 Praha 2
www.natur.cuni.cz

Department of Botany; Faculty of Science
Charles University in Prague
Faculty of Science
Benátská 2, 128 43 Praha 2
www.natur.cuni.cz/en

_______________________________________________
R-sig-ecology mailing list
[hidden email]
https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
Loading...