Fitting a mixture distribution to a variable subject to heteroscedastic measurement errors (Q1424624)

A mixture model with errors in variables is considered. The latent i.i.d. variables \(X_j\) have the mixture distribution \(\sum_{k=1}^m \alpha_k\phi_k(x)\), where \(\alpha_k\) are mixing probabilities and \(\phi_k\) are \(N(\vartheta_k,\zeta_k^2)\) densities. The observed variables are \(W_j=X_j+U_j\), \(j=1,\dots,N\), where \(U_j\) are independent \(N(0,\sigma_j^2)\) (\(\sigma_j\) are different for different \(j\) and known). An EM algorithm for approximate maximum likelihood estimation of \(\vartheta_k\), \(\zeta_k\) and \(\alpha_k\) is described. An estimator for the information matrix is derived. Results of a simulation study and analysis of Swiss data on Radon mean concentrations at different regions are presented.