[R-sig-ME] MCMCglmm and proportion time data
Jarrod Hadfield
j.hadfield at ed.ac.uk
Wed Mar 30 09:47:52 CEST 2016
Hi,
The Dirichlet distribution is probably most appropriate for a
continous-valued response, but I guess you could discretise it and treat
it as multinomial. I'm not sure how the results will change depending on
the scale at which you discretize - hopefully they will just go up by a
factor but I'm not sure. The family would be "multinomial4". The model
is set up like a multivariate model so you need to use trait to define
the different reposnes. Note that that there are only 3 traits, because
they are treated as contrasts from the base-line category (the first
response). The R-structure defined by us(trait):units is parameterised
by a 3x3 matrix, all elements of which are idenitifiable.
Alternativley you can define a base-line category (it doesn't matter
which, lets say STILL) and take the log ratio of each cetogory to the
base-line. This could then be treated as multivariate normal. The
problem then will dealing with the zeros. You may want to check out
Aitchinson's book on compsitional data.
Cheers,
Jarrod
On 17/03/2016 07:03, Marcus Michelangeli wrote:
> Hello,
>
> Apologies if a similar question has been asked and answered before, but any
> help would be extremely appreciated
>
> I measured 4 behavioral traits in the same test over a period of 20
> minutes. Each individual went through the test three times so I could
> estimate repeatability.
>
> EXAMPLE DATASET
>
> SKINK TRIAL POPULATION REGION ACTIVE BARRIER SHELTER STILL
>
> 1 SKINK22 1 Kuring-gai Chase N 1134.08 575.38 0.00
> 630.90
>
> 2 SKINK22 2 Kuring-gai Chase N 246.28 59.61 657.57
> 896.12
>
> 3 SKINK22 3 Kuring-gai Chase N 281.60 24.68 118.87
> 1381.20
>
> 4 SKINK23 1 Kuring-gai Chase N 0.00 0.00 1799.97
> 0.00
>
> 5 SKINK23 2 Kuring-gai Chase N 279.89 0.00 634.61
> 885.47
>
> 6 SKINK23 3 Kuring-gai Chase N 20.77 0.00 1243.55 535.65
>
>
>
>
> When looking at histograms of each behaviour, the data is heavily skewed to
> the left and is obviously non-normally distributed (all distributions look
> exponential)
>
> I'm not sure how to tackle this. I basically just want to estimate the
> repeatability of each behavioural trait and look at the effects of Region
> and Population on behaviour. I'm think my best option is to convert the
> data into proportions (i.e. what proportion of time did the individual
> spend doing each behavior over the 20 mins) and then run a GLMM (preferably
> using MCMCglmm in R or glmer) with a binomial distribution. My questions are
>
> 1. Would this be the best way to tackle this data?
> 2. What distribution should I use in MCMCglmm (is it family "categorical")?
> 3. Do binomial distributions require different priors. I usually just use
> an uninformative prior (V =1, nu =0.002)
>
>
>
> Thanks for your time
> Marcus
>
> [[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.
More information about the R-sig-mixed-models
mailing list