how to deal with multimodal likelihood
- From: "hyena" <as@xxxxxx>
- Date: Wed, 23 May 2007 10:36:40 +0200
Hi,
I am fitting a mixture discrete choice model using EM algorithm(kind of
iterative MLE), using different start values, I got several different
result, among, which, many of them looks not realistic(either too big
coeficients or too extreme group assignment, say, one groupwith 0.01 or less
proportion). I check with LIMDEP and found limdep result is more reasonable
in its estimated parameters and variances . after checking literatiures, I
am aware that this kind of mixture models are known for their multimodal
nature in likelihood function and advices is to deal with great care.
my questions is, how to find the global maxima or even just some reasonable
local maxima and make sure its parameters are not in a "spike" on the
likelihood surface? EM is guaranteed tofind a local maxima, so I need a way
to explore all or at least most of the modes. I am currently doing this
using random start values with lower and upper boundaries. Any better idea?
Another problem is the covergence of EM of painfully slow, I switch to a
direct search after EM improvement smaller then a certain value, will this
cause a problem?
Thanks in advance.
.
- Prev by Date: Re: Software to create 3-D scatterplot with regression plane
- Next by Date: NIPS*2007 - Final Call for Papers [Deadline: June 8, 2007, 23:59 Universal Standard Time]
- Previous by thread: Software to create 3-D scatterplot with regression plane
- Next by thread: NIPS*2007 - Final Call for Papers [Deadline: June 8, 2007, 23:59 Universal Standard Time]
- Index(es):
Relevant Pages
|
