Dear all,
It appears I sent a my request at the wrong time several weeks ago when many
were on holidays. I didn't receive a single response. Hope that the
non-response
doesn't indicate the difficulty of my problem. I am resending my request with
the hope that someone might assist.
------------------
Date: Tue, 15 Jun 1999 17:37:45 +0200
From: Sankoh <[log in to unmask]>
To: [log in to unmask]
Dear all!
I am a new WINBUGS user.
The new disposition models proposed by George E. Bonney (1998)
for correlated binary outcomes are based on the logistic regression
model. The proposed likelihood for m regions each of size n_{r}is given by:
L(\theta) = \prod_{r=1}^{m} \log
[(1-\alpha_{r})\prod_{j=1}{n_{r}}(1-y_{rj}) +
\alpha_{r}\prod_{j=1}^{n_{r}}d_{rj}^{y_{rj}}(1-d_{rj})^{1-y_{rj}}] (1)
where the parameters are as defined below. Note that when \alpha_{r}=1,
we get the standard logistic regression model (the term after the + in (1)).
Following Bayesian principles, we assume that \theta has a normal prior
(obvious reasons for a normal prior for the logistic regression since
there is no congugate prior for the logistic regression!) and build up the
log posterior likelihood function as in (2) below.
I am requesting assistant regarding the estimation of the the following
log posterior likelihood function denoted \log_{p}L(\theta) with the
bugs software:
\log_{p}L(\theta) =
\sum_{r=1}^{m} \log [(1-\alpha_{r})\prod_{j=1}{n_{r}}(1-y_{rj})+
\alpha_{r}\prod_{j=1}^{n_{r}}d_{rj}^{y_{rj}}(1-d_{rj})^{1-y_{rj}}]
- ||\theta - \mu||^{2}/2\sigma^{2} .............(2)
where y_{rj} is a binary response variable, r=1,...,m (regions),
j=1,...,n_{r} (individuals) with n_{r} the sample size of region r;
\theta = (\gamma_{0}, \gamma_{1}, ... ,\gamma_{k}; \lambda;
\beta_{1}, ... ,\beta_{s})
i.e. \theta \in R^{k+s+2}
\theta has a normal prior with mean = \mu and variance
\sigma^{2}I_{k+s+2},
Take: \mu = 0_{k+s+2}, \sigma^{2}=1;
e^{M(G_{r})+N(G_{r})+W(X_{rj})}
d_{rj} = ------------------------------------------
1+e^{M(G_{r})+N(G_{r})+W(X_{rj})}
M(G_{r}) = \gamma_{0}+\gamma_{1}G_{r1}+...+\gamma_{k}G_{rk}
N(G_{r}) = \lambda (constant, say 0.5)
W(X_{rj}) = \beta_{1}G_{1r1}+ ... +\beta_{sG_{srn_{r}}}
1+e^{-[M(G_{r})+N(G_{r})+(W(X_{rj})]}
\alpha_{r} = ----------------------------------------------
1+e^{-M(G_{r})}
Small sample size to test:
Health Status
of a tree labeling index tempertaure
Y Li T
REGION ONE
1 1.9 .996
1 1.4 .992
0 .8 .982
0 .7 .986
1 1.3 .980
REGION TWO
0 .6 .982
1 1.0 .992
0 1.9 1.020
0 .8 .990
0 .5 1.038
REGION THREE
1 1.0 1.002
0 1.6 .998
1 1.7 .990
1 .9 .986
0 .7 .986
Grateful for any assistance/tips - even if it is only to enter the
model (1) in BUGS.
With best wishes,
Osman Sankoh
Department of Statisitcs
University of Dortmund
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|