Abstract:
|
We propose an extension of the MCEM algorithm that exploits the assumption of unique parameterization, reduces bias due to imperfect data (missing and mismeasured data), and converges when implementation of the EM algorithm has a low convergence probability. Furthermore, the proposed extension reduces a potentially complex algorithm into a sequence of smaller, simpler, self-contained MCEM algorithms. Both theoretical and finite sample results are considered. Evidence suggests that partitioning can provide better bias reduction in the estimation of model parameters, particularly when there are parameters of primary and secondary interest. Breaking down a complex problem in to simpler steps may make the EM algorithm more accessible to a wider and more general audience permitting a broader implementation of the EM algorithm.
|