Nonparametric EM algorithms for estimating prior distributions (Q1179795): Difference between revisions

The author considers a sequence of repetitive experiments \(E_ 1,\dots,E_ N\), where each \(E_ i\) is represented by a pair of random vectors \((Y_ i,\theta_ i)\), where \(Y_ i\in R_ n\) are observed while \(\theta_ i\in \Theta\subset R_ p\) are not observed. Moreover, it is assumed that the \(Y_ i\)'s are independent but not identically distributed and the \(\theta_ i\)'s are independent identically distributed with unknown common distribution function \(F^*(\theta)\) defined on \(\Theta\). Let us denote \(p_ i(Y_ i\mid\theta_ i)\) the conditional density of \(Y_ i\) given \(\theta_ i\). Two alternate methods (for the discrete and the continuous case) to determine the maximum likelihood estimator of \(F^*\) which are motivated by the EM algorithm are developed. Both computational and convergence considerations are discussed and a nice illustrative example is presented. The main motivation for this problem comes from the need to estimate the distribution of parameters in pharmacokinetic models based on population data.