Navigation

Recent Publications

More...

Home

Selecting the number of classes in mixture regression models

u3002759 - Posted on 13 January 2015

Project Description: 

Due to its flexibility, mixture regression models have wide application in many areas such as survival analysis, longitudinal data, latent class model, classification and missing data. A major difficulty in building a mixture regression model is the unknown number of classes which makes the popular likelihood ratio test invalid to use. This project tries to study the performance of an proposed method to select the number of classes in mixture regression models and compare it with some existing methods in simulated and real data sets.

Researcher name: 
HUANG Beixue
Researcher position: 
Research Postgraduate Student
Researcher email: 
u3002759@connect.hku.hk
Research Project Details
Project Duration: 
01/2015 to 09/2015
Project Significance: 
In some preliminary work, the proposed method based on an measure of instability extended from clustering analysis shows better performance than some existing methods. It is expected that this method which is based on the geometry of the sample and considers the randomness, could give some improvements on the limitations of the current methods based on goodness of fit which may often be biased in real situations.
Remarks: 
Since this method needs large amounts of model fitting for random drawn samples at different stages which can be very time consuming especially for mixed models, HPC is very important in order to make comparison between this proposed method and the existing ones based on a large number of samples.