2003 Fiscal Year Final Research Report Summary
A Study on Non-ignorable Nonresponse Problems in Generalized Linear Models
| Project/Area Number |
12680310
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Statistical science
|
|---|
| Research Institution | University of Tsukuba |
|---|
Principal Investigator |
KANAZAWA Yuichiro University of Tsukuba, Institute of Policy and Planning Science, Associate Professor, 社会工学系, 助教授 (50233854)
|
|---|
| Project Period (FY) |
2000 – 2003
|
| Keywords | EM Algorithm / EM Algorithm / Non-ignorable nonresponse / Missing completely at random / Missing at random / Multiple imputation / Proportional-odds model / Logistic regression model / Structural equation model |
|---|
| Research Abstract |
Generalized linear models (GLM) with non-ignorably missing response or covariates has been studied since 1996 by Baker, Chen, Clifford, Fitzmaurice, Heath, Ibrahim, Laird, Lipshitz, Zahner. Their proposed method of estimation does not consider inestimability of the model parameters and their standard errors. In the fifth paper listed below, we uncovered the pattern of inestimability when some of the covariates in GLM are non-ignorably missing. In the fourth paper listed below, we again uncovered the pattern of inestimability when some of the resposes in GLM are non-ignorably missing. In this paper we also theoretically showed that if some of the responses are ignorably missing, then analysis of the listwisely deleted data will yield identical results with those using the method proposed by Ibrahim et al. in 1996 Biometrics paper. We also extended their parameter estimations method when there are non-ignorable non-responses in dependent variable to proportional-odds model and apply the
… More
method to total cast of ownership estimation in automobile markets in the U.S. in the paper listed third below
Structural equation models are widely recognized as an important tool for social sciences. The recent progresses in statistical software specializing in structural equation modeling such as LISREL, EQS, and AMOS accelerated its usage. In cases where each measurements are categorical, the model requires much more sophisticated approach and there does not seem to be an established consensus in how those categorical data ought to be analyzed. The recent advances in treating missing values in structural equation models are reviewed in Allison (2003) in which the limitations are pointed out of the more traditional methods such as listwise deletion or pairwise deletion and the more modern maximum likelihood estimation methods are explored. Also the underlying ideas and the actual implementation of multiple imputation methods are broadly discussed. This year we thoroughly reviewed the preceding progresses in this field and explore its problems in estimation using the real-world data in the first two papers listed below Less
|
