Penalized EM in Binomial Models with Missing Categorical Covariates

• Activity Number: 638
• Type: Contributed
• Date/Time: Thursday, August 5, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): John R. Stevens*+ and Jeremiah Rounds and David I. Schlipalius
• Companies: Utah State University and Purdue University and Queensland Government
• Keywords: Expectation-Maximization ; Firth's Penalized Maximum Likelihood ; Incomplete Data ; Rhyzopertha dominica ; Separation of Points

Abstract

In a dose-response study involving pesticide application, a categorical covariate was only recorded for surviving pests. In such cases, traditional maximum likelihood approaches alone cannot provide useful inference. An additional complication in these data is the possible separation of points, resulting in impractically large standard errors. We present an EM-based solution to missing categorical covariates in binomial models with logit links, addressing the separation of points by the use of a method of moments estimator for Firth's bias-reduction score-modification. We simulate to address questions about estimate bias, distribution, and appropriate coverage by confidence intervals.