The constrained Fisher scoring method for maximum likelihood computation of a nonparametric mixing distribution (Q2430223): Difference between revisions

The underlying distribution of an i.i.d. sample is modelled by a mixture model \(\int f(x,\vartheta)G(d\vartheta)\), where \(f\) is a known function, and \(G\) is an unknown mixing distribution. An iterative numerical algorithm is proposed for the computation of the nonparametric maximum likelihood estimate of \(G\). It is based on the fact that this estimate is always a discrete distribution: \(\hat G(\vartheta)=\sum_{j=1}^m \pi_j\delta_{\vartheta_j}(\vartheta)\). The constrained Fisher scoring method is used for computing the mixing probabilities \(\pi_j\) and the support points \(\vartheta_j\). Conditions for the convergence of the algorithm are given. Examples of real life and simulated data analyses are presented.