Standard errors of fitted component means of normal mixture (Q1298195): Difference between revisions

The model considered is the g-component normal mixture model with density function \(f(x,\psi)=\sum_{i=1}^g\pi_i\phi(x;\mu_i,\Sigma_i)\), where \(\phi(x;\mu_i,\Sigma_i)\) denotes the multivariate normal distribution density with unknown mean \(\mu_i\) and covariance matrix \(\Sigma_i\). The proportions \(\pi_1,\dots,\pi_g\) in which \(g\) components occur in the mixture are also unknown. The authors use maximum likelihood via EM algorithm to fit such a model and investigate the ways of calculating the standard errors of the estimated means \(\mu_i\), \(i=1,\dots,g\). Two methods of estimation of the standard errors are considered: the information method based on the inverse of the Fisher information matrix, and the computational intensive bootstrap method. Their performances are compared empirically by application to three real data sets and by a small-scale Monte-Carlo experiment.