This paper builds an approximate likelihood of a spatio-temporal model with missing data. The approximation is based on a version of the EM algorithm that uses marginal rather than conditional distributions at the E-step. It has a promise of being computationally more efficient due to reduction of the number of matrix inversions. The properties of this approximation are established analytically by using the techniques of estimating equations and matrix calculus. The estimates are shown to be biased, and corrections that restore consistency are proposed. An application to the EPA particulate matter data set is given.
A modification of the EM algorithm with applications to spatio-temporal modeling
Room
134