The multi-clump finite mixture distribution and model selection

Abstract

In practical data analysis, often an important problem is to determine the number of clumps in discrete data in the form of proportions. This can be done through model selection in a multi-clump finite mixture model. In this paper, we propose bootstrap likelihood ratio tests to test the fit of a multinomial model against the single clump finite mixture distribution and to determine the number of clumps in the data, that is, to select a model with appropriate number of clumps. Shortcomings of some traditional large sample procedures are also shown. Three datasets are analyzed.