Health Research and Policy

Abstract

DATE: January 31, 2013
TIME: 1:15 - 3:00 pm
LOCATION: Medical School Office Building, Rm x303
TITLE: Non-parametric empirical Bayes and Variance Estimation
SPEAKER: Marc A. Coram, Assistant Professor
Department of Health Research and Policy, Stanford

We describe a non-parametric empirical Bayes method that extends the work of Johns 1986. This method can be applied to matrix-shaped data in order to improve estimates of the mean, the variance, or other statistics for each row of the matrix. The assumption is that there is an underlying distribution for each row, such that the observations in that row are iid realizations from that distribution. The distributions themselves need not have any known form but should be iid realizations from some unknown population of distributions. The estimates are further improved by using auxilliary covariates with one realization per row. Cross-validation is employed to select an appropriately rich model.

Suggested reading:
M. Vernon Johns. Fully nonparametric empirical Bayes estimation via projection pursuit, Adaptive Statistical Procedures and Related Topics (John Van Ryzin, ed.), Institute of Mathematical Statistics, 1986, pp. 164--178.

Johns Jr, M. V. "Non-parametric empirical Bayes procedures." The Annals of Mathematical Statistics (1957): 649-669.

Stanford Medicine Resources:

Footer Links: