|
Activity Number:
|
369
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 6, 2008 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #301300 |
|
Title:
|
Constrained Maximum Likelihood Estimation for Logistic Regression
|
|
Author(s):
|
Martin Levy+ and ZhiYuan Dong*+ and Yan Yu and Tim Keyes
|
|
Companies:
|
Universtiy of Cincinnati and Universtiy of Cincinnati and Universtiy of Cincinnati and GE Capitol
|
|
Address:
|
Department of Quantitative Analysis, Cincinnati, OH, 452221, Department of Quantitative Analysis, Cincinnati, OH, 45221,
|
|
Keywords:
|
Linear Programming ; Optimization
|
|
Abstract:
|
Often prior information on event probabilities is available to logistic regression analysts. Classical maximum likelihood estimation ignores this information while Bayesian approaches can take full advantage of it. An intermediate approach couches the information in terms of constraints that either order the probabilities or set bounds on them. When these constraints translate into linear constraints on the estimated regression parameters, one can take computational advantage of common statistical packages to estimate them. Based on simulations we find that even with modest sample size and few constraints estimation accuracy is greatly improved over unconstrained MLEs based on a simple metric. Moreover, nice asymptotic properties of the constrained MLEs are retained. With bounding constraints, Bayesians can construe the bounds as noninformative priors on the event probabilities.
|
