Health Research and Policy

Abstract

DATE: September 26, 2013
TIME: 1:15 - 3:00 pm
LOCATION: Medical School Office Building, Rm x303
TITLE: Maximum Likelihood Multiple Imputation: A Review and a New Standard Error Formula
SPEAKER: Paul T. von Hippel
LBJ School of Public Affairs, University of Texas, Austin

Maximum likelihood multiple imputation (MLMI) is a form of imputation that imputes missing values conditionally on an asymptotically efficient parameter estimate, such as a maximum likelihood estimate. We contrast MLMI with the usual approach to imputation, which we call posterior draw multiple imputation (PDMI) because it imputes missing values conditionally on an parameter estimate drawn from the posterior. Both approaches have advantages. The advantages of MLMI are that it is simpler to implement and that it yields point estimates that are more efficient and less prone to small-sample bias. The advantage of PDMI is that it is compatible with a simple formula for estimating standard errors, whereas most formulas for MLMI standard errors are rather complicated. We propose and test a new formula for MLMI standard errors. The new formula is simple and works well provided the fraction of missing information is not too high. If the fraction of missing information is high, the new formula still works but requires a large number of imputations.

Suggested readings:
Paul T. von Hippel. Small-Sample Biases of Multiple Imputation and Maximum Likelihood in Incomplete Bivariate Normal Data. http://arxiv.org/abs/1307.5875

Stanford Medicine Resources:

Footer Links: