Bimodality of Plasma Glucose Distributions in Whites: A Bootstrap Approach to Testing Mixture Models

[摘要] The null distribution of the likelihood ratio test (LRT) of a onecomponent normal model versus two-component normal mixture model isunknown. In this paper, we take a bootstrap approach to the likelihood ratiotest for testing bimodality of plasma glucose concentrations from RanchoBernardo Diabetes Study. The small p-values from this approach support thehypothesis that a bimodal normal mixture model fits the data significantlybetter than a unimodal normal model. The size and power of the bootstrapbased LRT are evaluated through simulations. The results suggest that asample size of close to 500 would be necessary in order to attain a power of90% for detecting the unbalanced mixtures with means and variances similarto those in the Rancho Bernardo data. Besides sample size, the power alsodepends on the two means and variances of the two components in the data.

[学科分类] 土木及结构工程学

[关键词] EM algorithm;likelihood ratio test;mixture models