ABSTRACT
In 1977, Dempster, Laird, and Rubin (DLR) (Dempster et al., 1997) advocated a unified algorithm, called the EM algorithm, for deriving maximum likelihood estimates from incomplete data and showed its wide scope of application to various statistical models for the first time. Although many essays critical of its problems, such as the convergence in the algorithm, were released in the early days, numerous essays have been published over the past 20 years, hammering out new methodologies using the EM algorithm in almost all fields in which statistical analysis is required, including engineering, medical science, sociology, and business administration. According to a survey conducted by Meng and Pedlow (1992), at least 1700 essays exist on more than 1000 subjects. Moreover, Meng (1997) pointed out that more than 1000 essays were published in approximately 300 types of magazines in 1991 alone (of which statistics journals accounted for only 15%). These facts clearly indicate that the EM algorithm has already become a multipurpose tool for building a method ofstatistical analysis based on likelihood, surpassing the Newton-Raphson method and other substitution methods.
