Semiparametric Maximum Likelihood for Regression Models with Missing Data

• Activity Number: 294
• Type: Contributed
• Date/Time: Tuesday, July 31, 2007 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Suzanne Dubnicka*+
• Companies: Kansas State University
• Keywords: Horvitz-Thompson estimator ; kernel density estimation ; maximum likelihood ; missing at random

Abstract

Data sets from larger studies, such as AIDS Clinical Trials Group protocol 175 (ACTG 175), are often incomplete. Two methods which allow for the analysis of incomplete data are the EM algorithm and weighted estimating equations (WEE). We study the problem of estimating the parameters of a regression model when the error distribution is completely unknown and either the response variable or some of the covariates are missing at random. The proposed estimation procedure replaces the true unknown error density in the likelihood with a kernel density estimate and maximizes the estimated likelihood. The kernel density estimate is based on the Horvitz-Thompson estimator which also forms the basis for WEE. Thus, the proposed methodology shares common elements with both the EM algorithm and WEE. The method is illustrated on the ACTG 175 data and is verified via Monte Carlo simulation.