|
Activity Number:
|
61
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Sunday, August 3, 2008 : 4:00 PM to 5:50 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #302576 |
|
Title:
|
Algorithmic Errors in the Estimation of Tobit II Models and the Corresponding Failure To Recognize Selection Bias
|
|
Author(s):
|
Thomas Zuehlke and Anthony Kassekert*+
|
|
Companies:
|
Florida State University and Florida State University
|
|
Address:
|
655 Bellamy Building, Tallahassee, FL, 32306-2250,
|
|
Keywords:
|
Tobit II ; Heckman Selection Model ; Simultaneous Estimation ; Monte Carlo Simulation ; Software Review
|
|
Abstract:
|
Tobit II models (aka Heckman Selection models) are a standard statistical tool for detecting and correcting selection bias. ML estimation is complicated by the possibility of multiple roots to the score equations. Most software packages ignore this problem and may fail to converge to the global MLE even when consistent starting values are used. Convergence to the global MLE can be insured by use of a two-step algorithm which conducts a grid search over the bounded space of the error correlation, and then uses the conditional ML estimates as starting values for simultaneous estimation. The nature of the problem is illustrated using Monte Carlo simulation. Major software packages are then compared and found to suffer from the same algorithmic errors. Finally, replication of estimates for a sample of published data sets finds that roughly half of the studies report inaccurate estimates.
|
