On Single Versus Multiple Imputation for a Class of Stochastic Algorithms Estimating Maximum Likelihood
- 440 Downloads
We discuss a special class of stochastic versions of the EM algorithms. The advantage of the single imputation procedure in non-exponential family applications is highlighted. We prove that ergodic properties of the stochastic algorithms are dependent not on the multiplicity of the imputation scheme but rather on the stability of the deterministic component of an underlying stochastic difference equation.
KeywordsEM algorithm non-exponential family Markov chain stochastic difference equation
This paper is a modified version of a chapter from the author’s doctoral dissertation. The author thanks both Jean Diebolt and Ingram Olkin for their guidance, and for suggestions to this paper.
- Celeux, G. and Diebolt, J. (1985), The SEM algorithm: a probabilistic teacher algorithm derived from the EM algorithm for the mixture problem, Comp. Statist. Quart., 2, 73–82.Google Scholar
- Chan, K. (1990), Deterministic stability, stochastic stability and ergodicity, in: Tong, H., Non-linear Time Series (Oxford University Press, New York), pp. 448.Google Scholar
- Diebolt, J., and Ip, E.H. (1996), Stochastic EM: method and application, in: Gilks, W., Richardson, S., and Spiegelhalter, D. J., eds. Markov Chain Monte Carlo in Practice (Chapman and Hall, London) pp. 259–273.Google Scholar