[R-sig-ME] MCMCglmm error-in-variables (total least squares) model?

Dimitri Skandalis da.skandalis at gmail.com
Sun Jan 3 18:05:57 CET 2016


Hi Alberto,

Have you looked at the book Modern Phylogenetic Comparative Methods? R code
provided with Chapter 11 (2) deals with correlated measurements, and could
be a good place to start.

http://www.mpcm-evolution.org/practice/online-practical-material-chapter-11

Also, de Villemereuil et al. have developed an approach to related models
in BUGS/JAGS.

http://bmcevolbiol.biomedcentral.com/articles/10.1186/1471-2148-12-102

Dimitri

On Sun, Jan 3, 2016 at 8:16 AM, Jarrod Hadfield <j.hadfield at ed.ac.uk> wrote:

> Hi Alberto,
>
> When you say you have multiple observations for each species, do you mean
> that you have multiple observations for the response and the predictors? Do
> you expect the response and/or the predictors to be correlated at the
> observation level (for example are they measured on the same individuals)?
> I presume the answer to both these questions is yes if you wish to use the
> van de Pol method?
>
> Cheers,
>
> Jarrod
>
>
> Quoting Alberto Gallano <alberto.gc8 at gmail.com> on Sun, 3 Jan 2016
> 10:35:02 -0500:
>
> Hi Jarrod,
>>
>> I don't know the measurement error in the predictors in advance, so I
>> guess
>> it would need to be estimated simultaneously. I'm not 100% sure what you
>> mean by 'multiple observations for each predictor variable'. I have data
>> on
>> 132 species and have multiple observations (7 to 80) for each species. I'm
>> using a species level random effect and a phylogenetic covariance matrix
>> (using ginverse) to account for phylogenetic autocorrelation, and I'm also
>> using van de Pol and Wright's (2009) method for partitioning slopes into
>> between- and within-species (i'm interested in the between species slope).
>> My understanding is that neither of these things fits a model in which
>> orthogonal residuals are minimized.
>>
>> best,
>> Alberto
>>
>>
>> On Sun, Jan 3, 2016 at 5:24 AM, Jarrod Hadfield <j.hadfield at ed.ac.uk>
>> wrote:
>>
>> Hi Alberto,
>>>
>>> Do you know the measurement error in the predictors in advance or do you
>>> have multiple observations for each predictor variable and wish to
>>> estimate
>>> the error simultaneously?
>>>
>>> Cheers,
>>>
>>> Jarrod
>>>
>>>
>>>
>>>
>>>
>>> Quoting Malcolm Fairbrother <M.Fairbrother at bristol.ac.uk> on Sat, 2 Jan
>>> 2016 14:47:08 -0800:
>>>
>>> Dear Alberto (I believe),
>>>
>>>> To my knowledge, this is not possible in MCMCglmm (though Jarrod
>>>> Hadfield,
>>>> the package author, may weigh in with another response).
>>>> A collaborator and I have been working on a paper that shows how to fit
>>>> such models in JAGS (and perhaps Stan), though thus far we've only been
>>>> able to fit such models correcting for measurement error in the
>>>> predictors
>>>> at the lowest level. Multiple such predictors (including with different
>>>> measurement error variances) are no problem.
>>>> That paper, however, is probably still some months away from being
>>>> finished
>>>> and presentable. In the meantime, I don't know of any good options for
>>>> you.
>>>> If other subscribers to this list have any ideas, I'll be quite
>>>> interested
>>>> too!
>>>> - Malcolm
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> Date: Tue, 29 Dec 2015 16:09:53 -0500
>>>>
>>>> From: Alberto Gallano <alberto.gc8 at gmail.com>
>>>>> To: r-sig-mixed-models at r-project.org
>>>>> Subject: [R-sig-ME] MCMCglmm error-in-variables (total least squares)
>>>>>         model?
>>>>>
>>>>> I posted this question on Stack Overflow a week ago but received no
>>>>> answers:
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> http://stackoverflow.com/questions/34446618/bayesian-error-in-variables-total-least-squares-model-in-r-using-mcmcglmm
>>>>>
>>>>> This may be a more appropriate venue.
>>>>>
>>>>>
>>>>> I am fitting some Bayesian linear mixed models using the MCMCglmm
>>>>> package.
>>>>> My data includes predictors that are measured with error. I'd therefore
>>>>> like to build a model that takes this into account. My understanding is
>>>>> that a basic mixed effects model in MCMCglmm will minimize error only
>>>>> for
>>>>> the response variable (as in frequentist OLS regression). In other
>>>>> words,
>>>>> vertical errors will be minimized. Instead, I'd like to minimize errors
>>>>> orthogonal to the regression line/plane/hyperplane.
>>>>>
>>>>>    1. Is it possible to fit an error-in-variables (aka total least
>>>>> squares)
>>>>>    model using MCMCglmm or would I have to use JAGS / STAN to do this?
>>>>>    2. Is it possible to do this with multiple predictors in the same
>>>>> model
>>>>>    (I have some models with 3 or 4 predictors, each measured with
>>>>> error)?
>>>>>
>>>>>
>>>>>         [[alternative HTML version deleted]]
>>>>
>>>> _______________________________________________
>>>> R-sig-mixed-models at r-project.org mailing list
>>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>>
>>>>
>>>>
>>>>
>>>
>>> --
>>> The University of Edinburgh is a charitable body, registered in
>>> Scotland, with registration number SC005336.
>>>
>>>
>>>
>>>
>>
>
> --
> The University of Edinburgh is a charitable body, registered in
> Scotland, with registration number SC005336.
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>

	[[alternative HTML version deleted]]



More information about the R-sig-mixed-models mailing list