Date: Thu, 07 Jun 2012 10:31:40 +0200
From: "MORAN LOPEZ, TERESA" <tmoranlopez at mncn.csic.es>
To: Chris Howden <chris at trickysolutions.com.au>
Cc: "r-sig-mixed-models at r-project.org"
<r-sig-mixed-models at r-project.org>
Subject: Re: [R-sig-ME] Running multinomial models with random effects
Message-ID: <20120607103140.16232bmjomxwcx64 at webmail.csic.es>
Content-Type: text/plain
Thanks a lot Chris,
I will try that one and see what happens. However, if someone knows how to run multinomials with random effects I will be very greatful!
Quoting Chris Howden:
I'm not neccsarily advocating this. But another way is to use 3
logistic models with random effects. One for wether they choose big
acorn, one for small, one for both. Then compare the parameters of all
3 to see if there are any differences.
Chris Howden
Founding Partner
Tricky Solutions
Tricky Solutions 4 Tricky Problems
Evidence Based Strategic Development, IP Commercialisation and
Innovation, Data Analysis, Modelling and Training
(mobile) 0410 689 945
(fax / office)
chris at trickysolutions.com.au
Disclaimer: The information in this email and any attachments to it are
confidential and may contain legally privileged information. If you are not
the named or intended recipient, please delete this communication and
contact us immediately. Please note you are not authorised to copy,
use or disclose this communication or any attachments without our
consent. Although this email has been checked by anti-virus software,
there is a risk that email messages may be corrupted or infected by
viruses or other
interferences. No responsibility is accepted for such interference. Unless
expressly stated, the views of the writer are not those of the
company. Tricky Solutions always does our best to provide accurate
forecasts and analyses based on the data supplied, however it is
possible that some important predictors were not included in the data
sent to us. Information provided by us should not be solely relied
upon when making decisions and clients should use their own judgement.
On 07/06/2012, at 3:00, "MORAN LOPEZ, TERESA"
<tmoranlopez at mncn.csic.es> wrote:
Dear all,
I am a phD student working on animal behavior under different
predation risk. We have conducted an experiment in which acorn
selection by jays was evaluated in savanna vs forest-type
landscapes. We placed 8 feeders in two different forests and 8
feeders in two different savannas.
My response variable is factorial, acorn choice (big, small, both).
In my design I have spatial pseudorreplication within Areas. I have
not found multinom function which allows fitting random effects.
However, MCMCglmm package allows incorporating them using Bayesian
Methods. I am new using Bayesian Methods so it has taken me a lot of
effort to partially understand these models implementation. However,
once I have run my model I am not sure about output interpretation
and most importantly how to convert fix effect parameters into
probabilities.
I have tried three different things: to convert my data in binomial
(ignoring infrequent acorn choice ?both?) and run mixed model lmer.
Pros: I can include random effects , Cons: I lose an infrequent but
informative level of the response.
Keep categorical responses and run multinom function without random
effects. Pros: I can fit my response variable as categorical and
keep all levels. Cons: I am ignoring spatial autocorrelation
(however, when running binomial models acorn choice variation
between areas was much more lower than between risk levels).
+
Run MCMCglmm model with random effects. Pros: I can keep both
spatial autocorrelation and all levels of the response factor. Cons:
Not sure about model implementation and interpretation. Low sample
size.
Which option is the most accurate?
Thanks a lot! I am really stucked with my data.
Here I post my MCMCglmm model which is the one I have more doubts
#head(choice)-dataset with choice as factor (b=big, s=small, bt=both)
RISK Area CHOICE
LOW Sopie b
LOW Sopie b
LOW Sopie b
LOW Anchurones s
LOW Anchurones bt
LOW Anchurones b
I have read package tutorial and Hadfield notes (though I have not
read all the chapter of Hadfield course notes). Besides I have used
information posted Jaeger lab blog:
http://hlplab.wordpress.com/2009/05/07/multinomial-random-effects-models-in-r/
I am not very sure about random effects priors but I believe the
rest of them are ok.
k <- length(levels(choice$CHOICE))
I<- diag(k-1)
J <- matrix(rep(1, (k-1)^2), c(k-1, k-1))
#Constraints to variance-covariance matrix
prior = list(R = list(fix=1, V =(1/k) * (I + J), n=k), G = list(G1 =
list(V =diag(2), n=2)))
#R priors, Hadfield manual suggests that structure. Are they right?
NOT SURE ABOUT RANDOM PRIORS.
M<- MCMCglmm(CHOICE~ -1+ trait + RISK,
random = ~ us(trait):Area,
rcov = ~ us(trait):units,
prior = prior,
family = "categorical",
data = choice)
#-1+trait following Hadfield suggestions in order to estimate
interception if every outcome.
us(trait):----- since we are not sure of independence assumptions.
Is model function allright?
My model runs but I find it very difficult to interpret my summary
summary(M)
Iterations = 3001:12991
Thinning interval = 10
Sample size = 1000
DIC: 44.21802
G-structure: ~us(trait):Area
post.mean l-95% CI u-95% CI eff.samp
CHOICE.bt:CHOICE.bt.Area 14.622 0.1881 64.46 7.278
CHOICE.s:CHOICE.bt.Area 3.274 -11.1431 19.29 116.593
CHOICE.bt:CHOICE.s.Area 3.274 -11.1431 19.29 116.593
CHOICE.s:CHOICE.s.Area 5.071 0.1416 15.82 127.395
R-structure: ~us(trait):units
post.mean l-95% CI u-95% CI eff.samp
CHOICE.bt:CHOICE.bt.units 0.6667 0.6667 0.6667 0
CHOICE.s:CHOICE.bt.units 0.3333 0.3333 0.3333 0
CHOICE.bt:CHOICE.s.units 0.3333 0.3333 0.3333 0
CHOICE.s:CHOICE.s.units 0.6667 0.6667 0.6667 0
Location effects: CHOICE ~ -1 + trait + RISK
post.mean l-95% CI u-95% CI eff.samp pMCMC
traitCHOICE.bt -4.5042 -11.8926 0.3824 15.18 0.052 .
traitCHOICE.s -2.0187 -6.3370 1.3569 84.55 0.200
RISKRA?A 1.9095 -2.3541 8.2421 56.13 0.336
---
Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1
posterior.mode(M$Sol)
traitCHOICE.bt traitCHOICE.s RISKRA?A
-2.594132 -1.836527 1.554387
#Since there is not any intercept I can?t apply plogis() in order to
get probabilities of choice in different areas.
Sorry that I have that many questions and thanks a lot!!!