[quote]
引用第20楼hangover于2006-11-02 18:12发表的“”:
如果出现负数是很奇怪,好像问题出在那个"punishment" factor,忘了什么单词了,估计你能明白,对mixture model这个factor好像总是超常的大。可能是达到极限了, 所以才出现负数(猜的,没依据)。这个你要是弄出来,在这领域也小牛一下,呵呵,加油。
我很少用DIC,主要我是搞实用的,这个时候DIC好像拿不出手, 完全个人意见, 没别的意思,我的意思是投论文的时候。而且算chib's marginal likelihood或者Bayesian Factor也不是很麻烦,用这些别人也没话说。
priors我基本都用conjugate, 或者noninformative,省事, 问题就是很难检测conflicting information between likelihood and prior by simply looking at the posterior mean or posterior mode as the posterior is typically unimodal in conjugate framework. 当然如果postrior有两个kernel就很容易检测prior和likelihood谁分量大点。
.......[/quote]
我现在研究的是vairance component,所以non-informative prior对我来说不太可能。
Hierarchical model的DIC和pD不可信,这是我老板给我的例子:
----------------------------------------------<br />
model NormalDICCheck<br />
{<br />
mu ~ dnorm(0,0.000001)<br />
for(i in 1:10) { <br />
b[i] ~ dnorm(0,1)<br />
ymean[i] <- mu+b[i]<br />
y[i] ~ dnorm(ymean[i],1)<br />
}<br />
}<br />
mean sd MC_error val2.5pc median val97.5pc start sample<br />
mu 2.502 0.4472 7.561E-4 1.627 2.501 3.38 1001 1000000<br />
<br />
Dbar Dhat DIC pD <br />
y 26.5 21.0 32.0 5.498<br />
total 26.5 21.0 32.0 5.498<br />
<br />
model NormalDICCheck2<br />
{<br />
mu ~ dnorm(0,0.000001)<br />
for(i in 1:10) { <br />
y[i] ~ dnorm(mu,0.5)<br />
}<br />
}<br />
mean sd MC_error val2.5pc median val97.5pc start sample<br />
mu 2.499 0.4474 4.409E-4 1.621 2.5 3.376 1001 1000000<br />
Dbar Dhat DIC pD <br />
y 31.56 30.56 32.56 1.001<br />
total 31.56 30.56 32.56 1.001<br />
<br />
list(y=c(2,3,4,1,2,3,1,2,3,4)) <br />
<br />
----------------------------------
再介绍你在Bayesian Analysis的一篇文章:
G. Celeux¤, F. Forbesy, C.P. Robertz and D.M. Titteringtonx, 2006。 Deviance Information Criteria for Missing Data Models, Bayesian Analysis, 1, Number 4,
http://ba.stat.cmu.edu/forthcoming.php
第一篇。