Discussion:
[R-sig-ME] MCMCglmm: tissue type as a random effect when not all species have been measured for all tissuess
Nathanael Walker-Hale
2018-06-06 15:29:24 UTC
Permalink
Hi all,

I am planning to use MCMCglmm to do a phylogenetic comparative analysis. I
have multiple metabolite measurements on multiple tissue types per species
(e.g. leaf, three measurements, root, three measurements, per species). I
am interested in analyzing how levels of metabolite are predicted by the
presence or absence of a gene. Ideally, I would like to model both
between-species relationships (from a phylogeny) and tissue type as random
effects. However, not all species have measurements on all tissue types.

Will this be a problem for the analysis? Is it possible to run the model in
the presence of missing data like this? There is not a particularly heavy
bias to the pattern of missing tissue across the phylogeny, but some tissue
types have been measured much less than others (e.g. far fewer species have
floral measurements).

Best wishes,

Nathanael

[[alternative HTML version deleted]]
Ben Bolker
2018-06-06 15:46:42 UTC
Permalink
In principle this kind of missing data is no problem, but it seems odd
to treat tissue type as random if (as I understand it) you only have two
types of tissue (root, leaf) ... Perhaps you have more? Even so (say you
had a few more), this would present both technical issues [you'd
probably need a moderately strongly informative prior on the
among-tissue variance], and it seems odd to me to treat tissues as
exchangeable.
Post by Nathanael Walker-Hale
Hi all,
I am planning to use MCMCglmm to do a phylogenetic comparative analysis. I
have multiple metabolite measurements on multiple tissue types per species
(e.g. leaf, three measurements, root, three measurements, per species). I
am interested in analyzing how levels of metabolite are predicted by the
presence or absence of a gene. Ideally, I would like to model both
between-species relationships (from a phylogeny) and tissue type as random
effects. However, not all species have measurements on all tissue types.
Will this be a problem for the analysis? Is it possible to run the model in
the presence of missing data like this? There is not a particularly heavy
bias to the pattern of missing tissue across the phylogeny, but some tissue
types have been measured much less than others (e.g. far fewer species have
floral measurements).
Best wishes,
Nathanael
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
Ben Bolker
2018-06-06 17:09:06 UTC
Permalink
My guess is that it would be no problem to treat tissue as fixed and
species as random (and phylogenetically correlated) (and the
species:tissue interaction as random, so that you can model variation
within species -- **or** simply pool the multiple measurements from each
species/tissue combination, see Murtaugh "Simplicity and complexity in
ecological data analysis").
Why not try it (maybe with a subset of your data) and see what happens?
Hi Ben,
Thanks very much for your response. There are more tissues (six in all),
I just chose those as an example of the data format. I was under the
impression that the missing data would be a problem if attempting to
treat tissue type as fixed, but is this not the case? I have seen in
other threads the ways in which missing responses are imputed in
MCMCglmm, and I could perhaps merge tissue types into two or three
classes (say root, shoot and flower) and model them as fixed, if the
missing data does not create problems for the analysis.
Thanks again for your help,
Nathanael
  In principle this kind of missing data is no problem, but it seems odd
to treat tissue type as random if (as I understand it) you only have two
types of tissue (root, leaf) ... Perhaps you have more? Even so (say you
had a few more), this would present both technical issues [you'd
probably need a moderately strongly informative prior on the
among-tissue variance], and it seems odd to me to treat tissues as
exchangeable.
Post by Nathanael Walker-Hale
Hi all,
I am planning to use MCMCglmm to do a phylogenetic comparative
analysis. I
Post by Nathanael Walker-Hale
have multiple metabolite measurements on multiple tissue types per
species
Post by Nathanael Walker-Hale
(e.g. leaf, three measurements, root, three measurements, per
species). I
Post by Nathanael Walker-Hale
am interested in analyzing how levels of metabolite are predicted
by the
Post by Nathanael Walker-Hale
presence or absence of a gene. Ideally, I would like to model both
between-species relationships (from a phylogeny) and tissue type
as random
Post by Nathanael Walker-Hale
effects. However, not all species have measurements on all tissue
types.
Post by Nathanael Walker-Hale
Will this be a problem for the analysis? Is it possible to run the
model in
Post by Nathanael Walker-Hale
the presence of missing data like this? There is not a
particularly heavy
Post by Nathanael Walker-Hale
bias to the pattern of missing tissue across the phylogeny, but
some tissue
Post by Nathanael Walker-Hale
types have been measured much less than others (e.g. far fewer
species have
Post by Nathanael Walker-Hale
floral measurements).
Best wishes,
Nathanael
       [[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
<https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
<https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
Nathanael Walker-Hale
2018-06-06 17:29:56 UTC
Permalink
Thank you! I will.
Post by Ben Bolker
My guess is that it would be no problem to treat tissue as fixed and
species as random (and phylogenetically correlated) (and the
species:tissue interaction as random, so that you can model variation
within species -- **or** simply pool the multiple measurements from each
species/tissue combination, see Murtaugh "Simplicity and complexity in
ecological data analysis").
Why not try it (maybe with a subset of your data) and see what happens?
Hi Ben,
Thanks very much for your response. There are more tissues (six in all),
I just chose those as an example of the data format. I was under the
impression that the missing data would be a problem if attempting to
treat tissue type as fixed, but is this not the case? I have seen in
other threads the ways in which missing responses are imputed in
MCMCglmm, and I could perhaps merge tissue types into two or three
classes (say root, shoot and flower) and model them as fixed, if the
missing data does not create problems for the analysis.
Thanks again for your help,
Nathanael
In principle this kind of missing data is no problem, but it seems
odd
to treat tissue type as random if (as I understand it) you only have
two
types of tissue (root, leaf) ... Perhaps you have more? Even so (say
you
had a few more), this would present both technical issues [you'd
probably need a moderately strongly informative prior on the
among-tissue variance], and it seems odd to me to treat tissues as
exchangeable.
Post by Nathanael Walker-Hale
Hi all,
I am planning to use MCMCglmm to do a phylogenetic comparative
analysis. I
Post by Nathanael Walker-Hale
have multiple metabolite measurements on multiple tissue types per
species
Post by Nathanael Walker-Hale
(e.g. leaf, three measurements, root, three measurements, per
species). I
Post by Nathanael Walker-Hale
am interested in analyzing how levels of metabolite are predicted
by the
Post by Nathanael Walker-Hale
presence or absence of a gene. Ideally, I would like to model both
between-species relationships (from a phylogeny) and tissue type
as random
Post by Nathanael Walker-Hale
effects. However, not all species have measurements on all tissue
types.
Post by Nathanael Walker-Hale
Will this be a problem for the analysis? Is it possible to run the
model in
Post by Nathanael Walker-Hale
the presence of missing data like this? There is not a
particularly heavy
Post by Nathanael Walker-Hale
bias to the pattern of missing tissue across the phylogeny, but
some tissue
Post by Nathanael Walker-Hale
types have been measured much less than others (e.g. far fewer
species have
Post by Nathanael Walker-Hale
floral measurements).
Best wishes,
Nathanael
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
<https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
<https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
[[alternative HTML version deleted]]

Loading...