|
Activity Number:
|
205
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, July 30, 2007 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #309636 |
|
Title:
|
Improving the Efficiency of the Monte Carlo EM Algorithm Using Squared Iterative Methods
|
|
Author(s):
|
Ravi Varadhan*+ and Brian S. Caffo and Wolfgang Jank
|
|
Companies:
|
Johns Hopkins University and Johns Hopkins Bloomberg School of Public Health and University of Maryland
|
|
Address:
|
2024 E. Monument Street, suite 2-700, Baltimore, MD, 21205,
|
|
Keywords:
|
EM acceleration ; mixed models ; extrapolation methods ; SQUAREM
|
|
Abstract:
|
Implementation of the Monte Carlo EM (MCEM) algorithm is challenging mainly due to the difficulty in monitoring its convergence. Fixed sample size MCEM algorithms can be highly inefficient, since large MC samples are required to obtain even moderately accurate (2-3 significant digits) estimates. Caffo et al. (2005) have recently proposed an ascent-based MCEM algorithm that solves these problems. Here we evaluate whether the efficiency of the ascent-based MCEM algorithm can be further improved using a new class of numerical schemes called squared iterative methods (SQUAREM), which have recently been proposed to accelerate the convergence of the EM algorithm (Varadhan and Roland, 2004). SQUAREM can be easily implemented as it only requires the basic EM step. We present the results of two simulation examples: maximum likelihood (ML) estimation in multivariate-t and logit-normal models.
|
